NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

A Chinese Lab Put Its RNA Model in Nature — and Opened 21 Million Training Pairs

USTC's structRFM learns RNA structure alongside sequence, publishes in Nature and releases weights, code and data — a model with an unusually open posture.

2026-11-19 · 1078 words · NeuroAI
A Chinese Lab Put Its RNA Model in Nature — and Opened 21 Million Training Pairs

A strand of RNA is written in four letters. Read as text, it looks almost boringly simple. Folded, it becomes a shape that decides what a cell can do — and getting that shape wrong can change everything.

A team in Hefei has built a model that reads the letters and the folds at the same time. Two weeks ago it published the result in Nature and handed the weights to anyone who wants them.

On November 5, researchers at the University of Science and Technology of China (USTC) announced structRFM, an open-source foundation model (基础模型) for RNA sequences. The work appeared in Nature the same day. The team was led by Professor Zhou Shaohua (周少华) and tenured researcher Xiong Peng (熊鹏) of the university's MIRACLE lab; doctoral student Zhu Heqin (朱河勤) is first author. Unlike the chatbot large models (大模型) that dominate AI headlines, this one does not write prose. It predicts how a molecule folds and what it is likely to do.

Key takeaways

  • structRFM was trained on roughly 21 million computationally annotated pairs of RNA sequences and their secondary structures — a dataset the team built before training began.
  • The model takes a sequence as input and produces features at three levels: whole-sequence, individual nucleotide, and nucleotide-pair. Those features feed tasks including structure prediction and function inference.
  • Reported capabilities include zero-shot homology classification, improved secondary-structure prediction, splice-site prediction, internal ribosome entry site recognition and non-coding RNA classification.
  • The team released the training pairs, model weights and source code. That makes the result reproducible in a way that a closed model cannot be.

What the model actually learned

Most language models for biology treat RNA the way a text model treats a sentence: a long string of symbols whose meaning can be inferred from statistical patterns. That works surprisingly well, but it throws away something a biologist considers obvious. RNA is not a line. It folds back on itself, pairing bases into stems and loops, and that three-dimensional shape is tightly linked to function.

structRFM tries to keep both views. The team assembled about 21 million sequence-to-secondary-structure pairs, each computationally annotated, then designed an algorithm that learns sequence context and structural dependencies together during pretraining. After training, the model needs only a raw sequence. From that single input it can extract features describing the whole molecule, each nucleotide, and which nucleotides sit opposite each other.

Those three levels are what make the model useful across tasks. A sequence-level feature helps classify which family a molecule belongs to. Nucleotide-level features help locate splice sites — the points where a cell cuts and rejoins RNA. Pair-level features help predict which bases pair up, which in turn constrains the folded shape.

Why structure changes the problem

The team's claim is not that bigger is better. It is that adding the right kind of knowledge changes what a model can do with the same data. In ablation experiments — tests that remove one ingredient at a time — including structural information and a masking strategy improved performance. That is the kind of detail a press release usually skips and a researcher immediately looks for.

The reported results are broad rather than flashy. In zero-shot homology classification, where the model meets families it was not trained on, it performed well. Its features improved secondary-structure prediction, and when fed into a separate tertiary-structure prediction model, they improved the resulting three-dimensional shapes too. On function tasks it supported splice-site prediction, internal ribosome entry site recognition and non-coding RNA classification. Using a segmentation strategy, the team extended it to long non-coding RNAs.

For drug discovery, this matters because RNA is a target that has been hard to drug with conventional small molecules. A model that narrows the gap between a sequence and a plausible structure gives researchers a cheaper first pass before they spend months at a bench.

What is open — and what is not

Openness here is concrete. The lab says it has released the full paired dataset, the model weights and the source code. That is three different assets: data to retrain, weights to run, and code to audit. For a field where published results often cannot be reproduced because the training set stays private, releasing all three is a meaningful choice.

It is still a research release. There is no consumer product, no clinical claim and no evidence that the model replaces laboratory validation. The Nature paper is the evidence; the weights are the invitation.

The broader context is that Chinese research institutions have been publishing a run of high-profile science papers this year. This model is one entry in that run, notable for how little it withholds rather than for a record score.

Honest limitations

This article relies on Xinhua's wire report, China National Radio, the Chinese Academy of Sciences' reproduction of China Science Daily, and USTC's own news page. All describe the same announcement. I have not read the Nature paper in full, so the technical description follows the Chinese summaries; the DOI cited by the Academy of Sciences is 10.1038/s41586-026-11210-2, and readers should check the paper for exact figures and methodology.

"Improved performance" is the team's framing. The sources give no single benchmark table, no comparison against all competing RNA models, and no independent replication. Zero-shot results are promising but depend heavily on which families were held out. The 21 million training pairs are computationally annotated, which means the labels carry the limits of the annotation method.

I have also not verified the license attached to the released weights and code, so "open" here means publicly available, not necessarily free for every commercial use. Transliterations of names and the lab's English title may differ across outlets.

What readers can do now

  • If you work in computational biology, download the weights before deciding whether the claims hold for your data. A model that helps on one RNA family can fail on another; your own held-out set is the only test that matters.
  • If you are a student, treat the released dataset as the real prize. Twenty-one million annotated pairs is a resource even if you never run the model.
  • If you are simply following AI, notice the difference between a closed benchmark win and a reproducible result. The second is quieter and often more useful.

The model reads four letters and a fold. What the team released is the chance for anyone else to check whether it read them correctly — which, in science, is usually the point.

Related coverage

Cite this article

NeuroAI (脑机网), "A Chinese Lab Put Its RNA Model in Nature — and Opened 21 Million Training Pairs," 19 November 2026, https://neuroai.site/a/na-model-structrfm-rna-nature

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories