A Chinese Team Trained an RNA Model to Learn Shape, Not Just Sequence

Researchers at the University of Science and Technology of China have released an open RNA foundation model called structRFM, which they say improves predictions of how RNA molecules fold and what they do. The work was published in Nature on 5 November.

RNA is the molecule that carries instructions from DNA and, in many cases, does the work itself: building proteins, regulating genes, splicing messages. Its behaviour depends on shape. A single strand pairs with itself to form a secondary structure, which then folds into a three-dimensional one, and that shape decides what the molecule can bind to and how active it is.

The bottleneck has always been data rather than ideas. Sequencing has produced enormous numbers of RNA sequences, but experimental structures are scarce, so a model trained on sequence alone learns the alphabet without learning the grammar.

structRFM's approach is to put structure into the pretraining. The team built a dataset of about 21 million sequences paired with computationally annotated secondary structures, then trained the model to learn sequence context and structural dependence at the same time. Afterwards the model needs only a sequence as input, and produces features at three levels: whole sequence, individual nucleotide, and nucleotide pair.

In tests it classified RNA families it had never seen, improved secondary-structure prediction, and improved three-dimensional predictions when its features were plugged into existing folding models. It also handled functional tasks — predicting splice sites, identifying internal ribosome entry sites, and classifying non-coding RNAs — and held up better than comparable models on newly published RNA families.

The code, the model weights and the paired dataset have all been released. That matters more than usual here: in a field where structural data is the scarce resource, reproducibility is the difference between a benchmark number and a tool other laboratories can actually use.

The paper's first author is Zhu Heqin of the university's biomedical engineering school, with Zhou Shaohua and Xiong Peng as corresponding authors. It is the university's sixteenth paper in Nature, Science or Cell this year with the institution as first and corresponding author.