Back to news
research2026-11-05

USTC releases open-source RNA sequence foundation model structRFM on November 5 with Nature publication

A USTC MIRACLE lab team released structRFM, improving performance on RNA structure prediction and function inference downstream tasks, the 16th such publication for the university.

In the early hours of November 5, the University of Science and Technology of China announced that a team from its Medical Imaging Intelligence and Robotics Laboratory, led by Professor Zhou Shaohua and special researcher Xiong Peng, had released structRFM, an open-source RNA sequence foundation model that improves performance on downstream tasks including RNA structure prediction and function inference, with potential to support AI-driven scientific research, biological mechanism understanding and drug discovery. The results were published in the journal Nature, marking the university's 16th such first-author or corresponding-author publication in Nature, Science and Cell in 2026 so far.

RNA is a major biomolecule widely involved in protein synthesis, gene expression regulation and RNA splicing, and its function is closely tied to sequence and folded structure. RNA molecules form secondary structures through base pairing and fold further into three-dimensional conformations, and those structures shape biological activity and interactions with other molecules. As sequencing technology advances, researchers accumulate massive RNA sequence datasets; borrowing pretraining methods from natural language processing, RNA foundation models can learn sequence regularities from these datasets, extract general-purpose features usable across tasks, and support structure prediction and functional analysis.

structRFM was released in open-source form, with model weights and related evaluation code made public so researchers can reproduce results on structural biology tasks and iterate further. The team said the work demonstrates the practical role of pretrained foundation models in the scientific discovery pipeline, converting tasks once heavily dependent on human expertise into reusable model capability.

structRFMRNANature中国科大基础模型