An open structure-guided RNA foundation model for generalizable prediction

structRFM, a structure-guided RNA foundation model, is pre-trained on 21 million RNA sequence–structure pairs and ranks among leading models in zero-shot homology classification.

• Why it matters: RNA function depends on hierarchical structure, but prior RNA language-model pre-training often relies on noisy structural labels or task-specific objectives that can restrict generalization.
• What they did: The model integrates base-pairing into masked language modelling via a pair-matching operation, uses MUSES to reduce secondary-structure predictor bias, and applies a dynamic masking ratio.
• The result: structRFM sets new benchmarks for secondary-structure prediction, improves tertiary adaptation with Zfold, and performs comparably to AlphaFold3 across CASP15-natural, CASP16, and RNA-Puzzles.

AI summary of the abstract · may contain errors · not medical advice · read the paper.