Open RNA AI Model Rivals AlphaFold 3 at RNA Structure Prediction

A model that anyone can download now competes with the best closed system on RNA#

For protein folding, the question of whether artificial intelligence can match experiment has largely been settled. For RNA, it has not. RNA molecules fold into stems, loops and knotted shapes that decide what they do in a cell, yet predicting those shapes from sequence alone has stayed stubbornly hard. On 4 November 2026, a team including Heqin Zhu published a paper in Nature describing structRFM, an RNA foundation model that claims performance comparable to AlphaFold 3 on three RNA benchmarks (Zhu et al., Nature, 2026).

What makes the paper worth attention is not only the accuracy claim. The authors released the full 21-million-pair training dataset, the model weights and the source code, so other laboratories can test, criticise and extend the work (Zhu et al., Nature, 2026). That openness matters because it lets outsiders check the claims rather than take them on trust.

This article explains what structRFM is, how it differs from earlier RNA language models, what its results show, and what a careful reader should still ask. One caveat first: the full text of the Nature paper sits behind a subscription. Everything below comes from the open abstract, the figure titles, the extended data legends, the data and code availability statements, and the published bibliography. We say where that limits what we can conclude.

Background: the terms you need#

An RNA language model is a neural network trained on very large collections of RNA sequences. During training, some letters (A, C, G or U) are hidden and the model learns to guess them from context, a method called masked language modelling. The same idea powers modern text models, and it has been applied to proteins and to DNA. A foundation model is a large model trained once on broad data and then adapted to many specific tasks, such as classifying RNA families or predicting folds.

RNA secondary structure describes which bases pair with each other, forming stems and loops. Tertiary structure is the full three-dimensional shape. Secondary structure is cheaper to predict and is often the first step towards the 3D model. Scientists score predictions with several metrics. Root-mean-square deviation (RMSD) measures average atomic distance from the true structure, so lower is better. TM-score and GDT-TS measure overall similarity, so higher is better. INF scores how well base-pair interactions are recovered (Zhu et al., Nature, 2026).

The benchmarks matter too. CASP15 and CASP16 are community blind-prediction exercises, and RNA-Puzzles is a long-running set of RNA modelling challenges. In this paper the test sets contain 8 RNAs (CASP15 natural targets), 13 RNAs (CASP16) and 20 RNAs (RNA-Puzzles) (Zhu et al., Nature, 2026). Those numbers will come back later.

What structRFM is and what is new about it#

Earlier RNA language models learned from sequence alone. The Nature authors argue that this ignores the thing that shapes RNA function most, which is hierarchical structure (Zhu et al., Nature, 2026). A strong sequence-only comparator is RiNALMo, a 650-million-parameter model published in Nature Communications in 2025, which showed that general-purpose RNA language models can generalise well on structure prediction tasks (Penić et al., Nature Communications, 2025).

structRFM adds structure to pre-training. According to the abstract, it was trained on 21 million RNA sequence and structure pairs, and it builds base-pairing interactions into the masked language modelling step through a pair-matching operation (Zhu et al., Nature, 2026). In plain terms, the model is not only asked to fill in hidden letters. It is also pushed to learn which positions pair with which.

The model outputs three kinds of learned representation: sequence-level, nucleotide-level and pairwise features (Zhu et al., Nature, 2026). That design is what lets one pre-trained model be adapted to classification, secondary-structure prediction, 3D prediction and function prediction.

How the training data and annotations were handled#

Pairing information for 21 million sequences cannot come from experiments, because experimentally solved RNA structures number only in the thousands. The authors therefore needed predicted structures for most of the training set, and predicted structures carry the biases of whichever program made them. The paper's response is MUSES, a multi-source ensemble of secondary structures designed to reduce predictor bias (Zhu et al., Nature, 2026).

The extended data show that MUSES combines three existing predictors, BPfold, ViennaRNA and CONTRAfold, with weights tuned on a benchmark (Zhu et al., Nature, 2026). A second design choice is a dynamic masking ratio that balances structure-guided masking with ordinary nucleotide-level masking (Zhu et al., Nature, 2026).

The sequences themselves come from public resources. The authors list RNAcentral, Rfam, ArchiveII, RNAStrAlign, bpRNA and the Protein Data Bank among the sources for training and testing (Zhu et al., Nature, 2026). The pre-training dataset has its own Zenodo record, and the model code has a separate one (Zhu, Zenodo, 2026).

One feature of the data deserves a closer look. Ribosomal RNA makes up 66.3% of the pre-training set and transfer RNA 16.7%, leaving 17.0% for everything else (Zhu et al., Nature, 2026). Most sequences are around 250 nucleotides long, and canonical Watson-Crick and wobble pairs account for roughly half of all annotated base-pairing interactions (Zhu et al., Nature, 2026).

What the results claim#

The headline results fall into three groups. First, in zero-shot homology classification, where the model must group RNAs into families without any task-specific training, structRFM ranks highly (Zhu et al., Nature, 2026). Second, for secondary-structure prediction, the authors state that it sets new benchmarks (Zhu et al., Nature, 2026). Third, its 3D adaptation, Zfold, consistently improves three-dimensional predictions and the secondary structures extracted from them (Zhu et al., Nature, 2026).

Beyond structure, the model also improves on function-prediction tasks and generalises to newly released RNA families and long non-coding RNAs (Zhu et al., Nature, 2026). The extended data include internal ribosome entry site (IRES) identification and classification of non-coding RNA into 13 classes (Zhu et al., Nature, 2026).

The table draws on the abstract and the extended data legends (Zhu et al., Nature, 2026). We could not read the main-text tables, so exact scores are not reproduced here.

How to read "comparable to AlphaFold 3"#

AlphaFold 3 is a general model for predicting the structures of biomolecular complexes, including proteins, nucleic acids and small molecules (Abramson et al., Nature, 2024). The structRFM authors compared against it, against trRosettaRNA, and against a version of RiNALMo adapted for 3D prediction on the three RNA test sets (Zhu et al., Nature, 2026).

"Comparable" is a careful word, and it should be read as one. The abstract does not say structRFM beats AlphaFold 3, and a model that is specialised for RNA and open to inspection can be valuable even if it only ties a general system (Zhu et al., Nature, 2026). The more interesting claim is that the gain came from adding structure to pre-training, not from simply making the network bigger. That is a testable hypothesis, and because the code is public, others can now test it.

There is a statistical point as well. With 8, 13 and 20 targets, a handful of difficult RNAs can swing an average (Zhu et al., Nature, 2026). The authors present violin plots of RMSD for each target rather than only a mean, which helps readers see the spread (Zhu et al., Nature, 2026). Nature also published a peer review file alongside the paper, so the referees' concerns can be read directly (Zhu et al., Nature, 2026).

Limits the authors themselves report#

The extended data are frank about weaknesses. On CASP15, structRFM does better on natural targets than on synthetic targets with dense long-range interactions, and it sometimes misses short local stems, which leaves large multiway loops where there should be stems (Zhu et al., Nature, 2026). On CASP16, it does better on viral and aptamer targets and worse on multimer and bacterial targets (Zhu et al., Nature, 2026).

Sequence length is another problem. In the benchmark of six established secondary-structure methods on 3,975 independent sequences, performance dropped markedly as sequences grew longer, which the authors describe as a current challenge for long RNAs (Zhu et al., Nature, 2026). That limitation applies to the field's methods in general, and the legend does not claim structRFM escapes it.

The training mix is a further question. When two thirds of the data are ribosomal RNA, the model may be strongest on the kinds of RNA that are well represented and conserved across life (Zhu et al., Nature, 2026). The paper tests generalisation to newly released RNA families, which is the right kind of check, but independent replication on fresh, unusual RNAs would add confidence.

Finally, a few CASP16 targets used sequences that differed substantially from the experimental structures because unresolved regions were left unmodelled, so those comparisons are not clean (Zhu et al., Nature, 2026). None of these points is damning. They are the normal edges of a young field.

What this means for working biologists#

For a laboratory studying a non-coding RNA, a riboswitch or a viral RNA element, a free model that predicts secondary and tertiary structure is useful as a first hypothesis generator. It does not replace chemical probing, cryo-electron microscopy or crystallography. A predicted fold is a candidate to test, and the authors' own legends show that errors, such as missed stems, are visible on inspection (Zhu et al., Nature, 2026).

For computational groups, the open release is the main prize. The data, weights and code are on Zenodo and GitHub, so teams can retrain on new families, swap in better annotation ensembles, or test whether structure-guided pre-training helps for other molecules (Zhu, GitHub, 2026). The source code also has its own permanent archive (Zhu, Zenodo, 2026).

For the wider field of AI in bioinformatics, the paper is a reminder that scale is not the only lever. Adding biological structure to the training objective can help a model learn what sequence alone hides. Whether that holds as models grow, and whether it transfers to complex RNA-protein assemblies, remains open.

Frequently asked questions#

What is structRFM? It is an RNA foundation model trained on 21 million RNA sequence and structure pairs, which learns base-pairing information during masked language modelling (Zhu et al., Nature, 2026).

Does it beat AlphaFold 3 at RNA structure prediction? The abstract says its 3D adaptation, Zfold, performs comparably to AlphaFold 3 on CASP15 natural targets, CASP16 and RNA-Puzzles, not that it is better (Zhu et al., Nature, 2026).

Is the model free to use? The authors state that the dataset, model weights and source code are openly available (Zhu et al., Nature, 2026). The checkpoints and code are hosted on GitHub (Zhu, GitHub, 2026).

What is MUSES? MUSES is a multi-source ensemble of secondary structures that the authors designed to reduce the bias of any single prediction program (Zhu et al., Nature, 2026).

How large were the 3D test sets? They contained 8 RNAs for CASP15 natural targets, 13 for CASP16 and 20 for RNA-Puzzles, which are small samples (Zhu et al., Nature, 2026).

Where does it struggle? The authors report weaker results on synthetic targets with dense long-range contacts, and on CASP16 multimer and bacterial targets (Zhu et al., Nature, 2026).

How does it relate to RiNALMo? RiNALMo is an earlier sequence-based RNA language model with 650 million parameters, and structRFM was compared against an adapted version of it (Penić et al., Nature Communications, 2025).

Can I rely on its predictions without experiments? No. Treat a predicted fold as a hypothesis, because the paper's own legends show missed stems and weaker performance on some target types (Zhu et al., Nature, 2026).

References#

1. Zhu, H., Li, R., Chang, A. et al. An open structure-guided RNA foundation model for generalizable prediction. Nature (2026). https://www.nature.com/articles/s41586-026-11210-2
2. Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493-500 (2024). https://doi.org/10.1038/s41586-024-07487-w
3. Penić, R. et al. RiNALMo: general-purpose RNA language models can generalize well on structure prediction tasks. Nature Communications (2025). https://www.nature.com/articles/s41467-025-60872-5
4. Zhu, H. structRFM pre-training data. Zenodo (2026). https://doi.org/10.5281/zenodo.18272040
5. Zhu, H. structRFM source code. Zenodo (2026). https://doi.org/10.5281/zenodo.18274292
6. Zhu, H. structRFM repository. GitHub. https://github.com/heqin-zhu/structRFM