Chinese scientists launch open-source RNA model structRFM, published in Nature

The team from the University of Science and Technology of China has launched structRFM, an open-source foundational model for RNA sequences, trained on 21 million sequence-structure pairs, achieving cutting-edge performance on multiple core metrics and published in Nature.

Author and source: AIBase

A research team from the University of Science and Technology of China has recently achieved a major breakthrough at the intersection of life sciences and artificial intelligence, officially launching structRFM, an open-source foundational model for RNA sequences. The related research has been published in the prestigious international journal Nature. To advance the global scientific ecosystem, the team has fully open-sourced the complete paired data, model weights, and source code of structRFM, making them freely available to researchers worldwide with no restrictions.

As a crucial class of macromolecules within living organisms, ribonucleic acid (RNA) plays a widespread and regulatory role throughout the entire lifecycle of gene expression and protein synthesis, making it an essential target in global innovative drug development.

At the microscopic level, RNA molecules form complex secondary structures through base pairing and further fold into specific three-dimensional conformations. It is these three-dimensional and secondary structures that directly determine the molecule’s biological activity and its interactions with other molecules. However, precisely deciphering RNA structure and function has long been a challenge in the field of bioinformatics.

To tackle this challenge, the research team from the University of Science and Technology of China took an innovative approach, constructing approximately 21 million precisely annotated sequence-secondary structure pairs, providing a vast and high-quality data foundation for training large models.

On this foundation, the research team meticulously designed an innovative algorithm that enables the model to simultaneously learn sequential context and structural dependencies during pre-training. After extensive pre-training, the model demonstrates exceptional convenience and generalization capability: researchers need only input a basic sequence, and the model automatically and accurately extracts triple-level features—at the sequence, nucleotide, and nucleotide-pair levels—efficiently supporting a variety of structure and function inference tasks.

In rigorous experiments and benchmark tests, structRFM demonstrated outstanding all-around performance:

Zero-shot classification and structure prediction: The model excels in zero-shot homology classification tasks and significantly improves the accuracy of RNA secondary structure prediction.

Three-dimensional conformation co-optimization: When researchers seamlessly integrated their extracted features into the 3D structure prediction model, the final predictions of the 3D conformation and its corresponding secondary structure were significantly improved.

Precise inference of biological functions: In practical functional inference applications, the model excels at splice site prediction, internal ribosome entry site (IRES) identification, and non-coding RNA classification.

Strong generalization and scalability: Even when faced with newly discovered RNA families, the model maintained excellent performance in secondary structure prediction; through a flexible segmentation strategy, it successfully extended its application to the classification of long non-coding RNAs.

In addition, the research team conducted meticulous ablation studies to scientifically validate the critical role of incorporating structural information and masking strategies in enhancing the model’s overall performance, providing robust experimental support for future algorithm design. With the full open-sourcing of this model, global scientific communities can leverage this new tool to accelerate progress in RNA structure analysis and cutting-edge biomedical research.