AI-generated concept illustration: a folded single-strand ribbon and connecting threads evoke RNA folding and base pairing
An RNA sequence is a string of characters, but when the molecule performs its function, the relationships among those characters are not limited to adjacency. Bases that are far apart can pair, and the sequence thereby forms a folded structure. A recently published study introduced by the University of Science and Technology of China brings this pairing relationship into foundation model training, so that when the model restores masked sequence positions, it also learns structural connections.
According to a USTC announcement on November 5, the paper by Zhou Shaohua and Xiong Peng’s team, “An open structure-guided RNA foundation model for generalizable prediction,” was published online in Nature in the early hours of that day. The model is named structRFM, and the research focuses on RNA structure prediction and functional inference. This round treats the paper publication as the new event; the authors’ repository still lists a 2025 preprint, so this report should not be written as the project’s first appearance or first open-sourcing that day.[1][2]
According to the university, the team used MUSES, a multi-source secondary-structure integration method, to integrate results from different prediction methods and construct paired data of about 21 million sequences and secondary structures. These structures were computationally annotated, which does not mean that all 21 million have experimentally determined structures. Model training uses both structure guidance and random masking, enabling the sequence-recovery task to use pairing information without relying entirely on one fixed masking approach.[1]
The study’s evaluation covers two categories of problems, structural and functional, including secondary structure prediction, providing features for tertiary structure prediction, and tasks such as splice site prediction, internal ribosome entry site prediction, and noncoding RNA classification. The university says the relevant performance improved; these are experimental results disclosed by the research team, and this round did not rerun the model, nor does it generalize them as advantages that hold for all RNA problems.[1]
Methodologically, the key question raised by this work is how foundation models use imperfect structural information. Computational prediction can supplement the many sequences lacking experimental annotation, but predictions themselves contain errors. Combining information from multiple sources is the approach adopted in the study; external review still needs to examine performance under different data splits and RNA families to determine which improvements can transfer to new tasks. This is editorial analysis of the research method.
The public repository offers a more concrete reuse path than a single online demo: researchers can consult the source code, pretrained model entry points, and instructions for downloading sequence-structure data, and can use feature extraction or fine-tuning workflows according to the task. The repository provides an MIT license file, but in practice users should still separately check the descriptions of the downloaded data and other dependencies. This round confirmed these public entry points but did not download all files one by one or verify the training workflow.[2]
The value of open resources lies in giving external teams the opportunity to use the same set of materials to probe the results. For example, whether performance is maintained when structural annotations come from a different source; when facing families rarely seen in training, whether errors concentrate in a certain type of pairing; whether the strategy for handling long sequences loses cross-segment relationships. These questions require follow-up experiments to answer and cannot be replaced by “the model is open source.”
For readers interested in AI research tools, structRFM’s progress illustrates the importance of training-task design: not only expanding data volume, but also changing what the model is required to learn from the data. The paper publication and resource release provide a starting point for further scrutiny; applications such as drug discovery remain directions for follow-up research, and current evidence does not support describing it as a delivered drug outcome or clinical capability.
Sources:
