Does structure-guided AI improve RNA prediction? An open 21-million-pair model led several benchmarks—but biology still needs experiments
Nature's structRFM pairs RNA sequences with predicted secondary structures during pre-training and releases its data, weights and code. It improved several secondary-, tertiary- and function-prediction benchmarks, but the structural labels are computational, long RNAs remain difficult and benchmark accuracy is not a therapeutic result.
The short version
- What happened
- structRFM was pre-trained on 21,477,078 RNA sequence–structure pairs, but the structures were generated by another predictor rather than measured experimentally.
- Who it affects
- Researchers can reuse open data and weights rather than repeat expensive pre-training, broadening access to RNA modelling.
- What remains unproven
- The 21 million structural labels were generated computationally by BPfold rather than measured in laboratories.
Keep in mind
The 21 million structural labels were generated computationally by BPfold rather than measured in laboratories.
The encoder accepts sequences up to 512 nucleotides and performed much worse on four long synthetic CASP15 targets.
Some tertiary-structure comparisons use only eight natural CASP15 and 20 RNA-Puzzles targets.
Reported performance belongs to task-specific pipelines and selected metrics, not the encoder alone or real-world therapeutic outcomes.
Multi-source analysis · 2 linked sources. Primary analysed source date: 4 November 2026. Earliest linked supporting source: 6 August 2025.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 8 min The evidence, methods, implications and limitations
The direct finding: structural guidance improved several RNA benchmarks
A peer-reviewed Nature paper published on 4 November presents structRFM, an open RNA foundation model designed to learn both nucleotide sequence and base-pair structure. The team reports leading or near-leading results across zero-shot homology, secondary-structure, tertiary-structure and function-prediction tasks. On two fine-tuned secondary-structure tests, structRFM reached F1 scores of 0.873 on 3,611 ArchiveII600 RNAs and 0.641 on 1,305 bpRNA-TS0 RNAs. Removing the structure-guided pre-training reduced performance by 2.7% and 3.2% respectively, providing an internal ablation rather than only a leaderboard comparison.
For tertiary structure, the derived Zfold system reported root-mean-square deviation of 14.8 angstroms on eight natural CASP15 targets and 6.8 angstroms on 20 RNA-Puzzles targets. The paper compares those values with AlphaFold3's 15.6 and 8.1 angstroms, improvements of about 5.4% and 19.1%. Lower is better, but the denominators are small and benchmark composition matters. On four synthetic CASP15 RNAs, structRFM's mean RMSD was 42.9 angstroms—far worse than on natural targets. The responsible conclusion is that structure-guided pre-training strengthened several benchmark pipelines, not that RNA structure is solved.[1][2]
The training denominator is large, but its structures are predictions
Researchers filtered 21,477,078 active RNAcentral sequences no longer than 512 nucleotides. They split 20,403,225 into training and 1,073,855 into a held-out test set. Because experimental structures do not exist at that scale, the team used BPfold, a separate secondary-structure predictor, to generate base-pair labels and confidence information. The resulting corpus is useful and publicly released, but it is not 21 million laboratory measurements. Systematic errors in BPfold can become pre-training priors, especially for non-canonical pairings, dynamic conformations and long-range interactions that the label generator does not capture well.
Pre-training ran for 3,238,154 steps with batches of 64 on two NVIDIA A100 80GB GPUs for about 400 hours. The backbone is a 12-layer BERT-style transformer with hidden dimension 768 and a maximum sequence length of 512. That makes the model far smaller than many general-purpose language models and easier for research groups to inspect or adapt, yet the compute and hardware are still substantial. Openness improves reproducibility; it does not remove distribution bias from RNAcentral, predicted labels or filtering choices. Users should preserve the distinction between an open resource and an experimentally validated biological atlas.[1][2]
The distinction that matters
Where the structure information enters—and where experiments remain essential
structRFM learns from predicted base pairs, transfers representations to benchmark tasks and still requires laboratory confirmation.
Source sequences
21,477,078 RNAcentral entries
Large but database-shaped corpus
Predicted structure
BPfold supplies base-pair labels
Scalable guidance with inherited error risk
Pre-training
Mask nucleotides and paired partners
Joint sequence–structure representations
Benchmarks
Fine-tune task-specific pipelines
Gains vary by dataset and metric
Biology
Validate molecules experimentally
Needed before therapeutic claims
How structure-guided masking changes the learning task
Ordinary masked language modelling hides random nucleotides and asks a model to reconstruct them. structRFM adds a PairMatch operation: when a masked nucleotide participates in a canonical base pair, its paired partner can also be masked. This makes the reconstruction task depend on surrounding sequence and the relationship between paired positions, encouraging the network to encode more than local nucleotide frequency. A schedule gradually increases the share of structure-guided masks from zero to half of the masking process, while retaining standard masks so the model learns both sequence-level and pair-level patterns.
After pre-training, structRFM produces a sequence classification representation, nucleotide-level embeddings and a pairwise matrix representation. Task-specific heads or existing structure modules then turn those features into predictions. Secondary-structure experiments attached MXfold2 and dynamic-programming post-processing; tertiary-structure experiments replaced trRosettaRNA's predicted contact map with structRFM features before PyRosetta refinement. This modular design helps explain why results can improve without one monolithic model doing everything. It also means headline performance belongs to a pipeline that includes task-specific training, multiple-sequence alignments and established post-processing—not to the pre-trained encoder alone.[1][2]
The evaluation spans many tasks, but no one metric proves general biological understanding
Zero-shot homology tests sampled 100,000 homologous and 100,000 non-homologous pairs from both Rfam and ArchiveII. structRFM ranked fourth among 15 biological language models on the reported overlap measure. A separate zero-shot structure test used only 116 PDB RNAs and achieved a maximum F1 of 0.44 at a threshold selected from a dense search, showing moderate rather than decisive unsupervised structure recovery. Fine-tuning then improved task performance, as expected, but those downstream datasets bring their own labels and possible overlaps with the biological families represented during pre-training.
Function tests included 16,505 held-out splice-site sequences, 1,164 balanced IRES test sequences and a 13-class non-coding RNA classification task. structRFM reached top-k accuracy 0.42 on the SpliceAI benchmark—behind the specialized SpliceBERT—and IRES AUROC 0.6972 with F1 0.708. These are useful results, not clinical-grade discrimination. The paper's breadth is a strength because it reduces the risk that the method was built for one narrow leaderboard. At the same time, multiple benchmarks create many possible headline numbers. Readers should examine the task-specific denominator, comparator and operating metric instead of treating 'state of the art' as one transferable property.[1][2]
Open release increases scientific value, while the main limits remain biological
The authors released the 21-million-pair corpus, pre-trained and fine-tuned checkpoints, code and source data. That allows other teams to audit splits, test family leakage, replace predicted structures and evaluate unfamiliar RNA classes. It also lowers the barrier for laboratories that could not afford to reproduce 400 hours of A100 pre-training. The 2025 preprint remains part of the evidence history; the current news is peer-reviewed publication, not a brand-new dataset created this week. Exact model cards, licenses and downstream validation requirements still matter for anyone adapting the system.
The model's 512-nucleotide limit is consequential. Long non-coding RNAs and synthetic constructs can contain distant interactions beyond that window, and the paper reports weaker performance on long synthetic CASP15 targets. Secondary structure is dynamic, environment-dependent and includes non-canonical interactions that a single predicted label cannot fully represent. Root-mean-square deviation also compresses a three-dimensional prediction into one average distance and may not capture whether a biologically critical pocket or motif is correct. Laboratory probing, cryo-electron microscopy, crystallography and functional assays remain the arbiters of what a molecule actually does.[1][2]
What changes for researchers—and what would change this assessment
For computational biologists, structRFM provides an unusually complete starting point for studying whether structural priors improve RNA models. For experimental teams, it may help prioritize candidates, design measurements or flag likely structures, reducing search costs without replacing experiments. Claims about RNA medicines require much more: prospective design, synthesis, stability, cellular activity, toxicity, delivery and clinical testing. No patient received a treatment in this study, and no drug candidate was shown to work because of the model. The practical impact today is a research tool and open benchmark resource.
Confidence would rise with leakage-resistant family splits, independent replication from the released weights, experimental rather than predicted pre-training structures, uncertainty calibration and prospective tests on RNAs deposited after model development. Comparisons should include compute-matched baselines and report target-level failures, not only averages. The assessment would weaken if gains disappear after removing close homologues or if predicted structural labels simply transfer BPfold's biases. This is consequential peer-reviewed science but not Breaking: it advances computational RNA modelling without reporting an urgent therapy, regulatory action or public-health event.[1][2]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring drug discovery: compare the finding and its evidence limits.
An AI scored 1.1 million rat sleep segments—but how far does 13 animals take us?
A peer-reviewed model retained high sleep-stage accuracy across three serotonergic drug conditions. The subject-held-out design is useful, but all recordings came from 13 male rats in one laboratory.
Read the next reportExplore the full source and evidence record
What this means for people
- Researchers can reuse open data and weights rather than repeat expensive pre-training, broadening access to RNA modelling.
- Experimental laboratories may prioritize candidates more efficiently, but must verify predicted structures and functions.
- Patients should not interpret benchmark gains as evidence that an RNA medicine is effective or safe; no therapy was tested.
Global context
RNA foundation models could support basic biology, diagnostics and therapeutic research, but training data and benchmark families are uneven. structRFM's open release offers researchers worldwide a resource that would otherwise require substantial compute. The core scientific question is whether its structural priors generalize to newly measured, evolutionarily distant and engineered RNAs. That question is more important than winning familiar benchmarks and requires independent experimental collaboration.
What the evidence does not yet show
- The 21 million structural labels were generated computationally by BPfold rather than measured in laboratories.
- The encoder accepts sequences up to 512 nucleotides and performed much worse on four long synthetic CASP15 targets.
- Some tertiary-structure comparisons use only eight natural CASP15 and 20 RNA-Puzzles targets.
- Reported performance belongs to task-specific pipelines and selected metrics, not the encoder alone or real-world therapeutic outcomes.
What to watch next
- Leakage-resistant evaluation on newly released RNA families.
- Independent replication from public weights and code.
- Training with experimental, dynamic and non-canonical structural data.
- Prospective laboratory studies of model-prioritized molecules.
Living evidence record
Impact record IAI-1NHJMEO
Evidence stage
Studied
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
5 November 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 5 November 2026
Research question, topics and regions
Whether pairing RNA sequences with predicted base-pair structure during foundation-model pre-training improves generalization across RNA structural and functional prediction tasks
Topics: RNA · Foundation models · Drug discovery · Biology · Open science · Structure prediction
Regions: China · Global
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.