A research team from the University of Science and Technology of China (USTC) has recently achieved a major breakthrough at the intersection of life sciences and artificial intelligence, officially launching an open-source RNA sequence foundation model named structRFM. The related research findings were recently published in Nature, a top international academic journal. To promote the shared progress of the global research ecosystem, the team announced that it has fully open-sourced the model's complete paired data, model weights, and source code, making them freely available to researchers worldwide without barriers.
Background: Tackling the Underlying Bottlenecks in RNA Research and New Drug Development
As a critical core macromolecule in living organisms, ribonucleic acid (RNA) is widely involved in and regulates the entire life cycle of gene expression and protein synthesis, making it an indispensable key target in global innovative drug development today.
At the microscopic level, RNA molecules form complex secondary structures through base pairing and further fold into specific three-dimensional conformations. These 3D and secondary structures directly determine the biological activity of the molecules and how they interact with other molecules. However, accurately resolving the structure and function of RNA has long been a major challenge in bioinformatics.
Architecture and Data: Dual Pre-training Driven by 21 Million Paired Records
To overcome this challenge, the USTC research team took an innovative approach and constructed approximately 21 million paired sequence-secondary structure data records with precise computational annotations, laying a massive and high-quality data foundation for training the large model.
On this basis, the R&D team carefully designed an innovative algorithm that enables the model to simultaneously learn sequence context information and structural dependencies during pre-training. After deep pre-training, the model demonstrated exceptional ease of use and generalization ability: researchers only need to input the basic sequence, and the model can automatically and accurately extract three levels of features—sequence-level, nucleotide-level, and nucleotide-pair-level—efficiently supporting a variety of structure and function inference tasks.
Experimental Results: Multiple Core Metrics Reach International Cutting-edge Levels
In rigorous experiments and benchmark tests, structRFM demonstrated impressive all-around performance:
Zero-shot classification and structure prediction: The model performed excellently in zero-shot homology classification tasks and significantly improved the accuracy of RNA secondary structure prediction.
Collaborative optimization of 3D conformation: When researchers seamlessly integrated the features extracted by the model into a 3D structure prediction model, the final predictions of the 3D conformation and its corresponding secondary structure were significantly improved.
Accurate inference of biological function: In practical functional inference applications, the model can excellently support splice site prediction, internal ribosome entry site (IRES) identification, and non-coding RNA classification.
Robust generalization and extensibility: When faced with newly discovered RNA families, the model still maintained excellent secondary structure prediction performance; through a flexible segmentation strategy, it also successfully extended its application boundary to long non-coding RNA classification tasks.
In addition, through detailed ablation experiments, the research team further scientifically verified the key roles of introducing structural information and the masking strategy in improving the model's overall performance, providing solid experimental data support for future algorithm design. With the full open-sourcing of the model, global research efforts will be able to rely on this new tool to accelerate RNA structure resolution and the development of cutting-edge biomedicine.
