An open structure-guided RNA foundation model for generalizable prediction.

structRFM, trained on 21 million RNA sequence–structure pairs, learns joint sequence and structure representations that rank highly in zero-shot homology classification and set new secondary-structure prediction benchmarks.

• 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, limiting generalization across diverse RNA tasks and families.
• What they did: The model integrates base-pairing interactions into masked language modeling via a pair-matching operation, using MUSES multi-source secondary structures and a dynamic masking ratio to balance structure-guided and nucleotide masking.
• The result: structRFM improves tertiary-structure adaptation with Zfold, consistently boosting 3D predictions and extracted secondary structures with performance comparable to AlphaFold3 1 across CASP15-natural, CASP16, and RNA-Puzzles, and it also improves function prediction and generalizes to newly released RNA families and long non-coding RNAs.

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