What it is
RNA language models perform well across many tasks by learning from large sequence collections, but an RNA's function is shaped by its hierarchical structure, so the authors built structure into pre-training. structRFM is trained on 21 million RNA sequence-structure pairs, folding base-pairing interactions into masked language modelling through a pair-matching operation. The authors also introduce MUSES, a multi-source ensemble of secondary structures meant to reduce the bias of any one structure predictor, and a dynamic masking ratio that balances structure-guided and nucleotide-level masking. structRFM ranks highly in zero-shot homology classification and establishes new benchmarks for secondary-structure prediction. Its tertiary-structure adaptation, Zfold, consistently improves three-dimensional predictions and the secondary structures extracted from them, with performance comparable to AlphaFold3 across the CASP15-natural, CASP16 and RNA-Puzzles benchmarks.
Why it matters
structRFM shows that building base pairing into pre-training gives an RNA model that carries over to structure and function tasks. The authors report that it also improves performance on function-prediction tasks and generalizes to newly released RNA families and long non-coding RNAs, a test of whether it learned more than the families it was trained on. The complete 21-million-pair dataset, model weights and source code are openly available, so other groups can build on it.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed underRNA and protein synthesis mechanisms, Machine Learning in Bioinformatics, Protein Structure and Dynamics