The abstract is read at the publisher; the summary is JClub's.
RNAtomy, a self-supervised RNA foundation model, learns inter-nucleotide geometry and residue conformations and achieves particularly strong contact and long-range base-pair prediction.
- Why it matters: Most RNA foundation models rely mainly on sequence or secondary structure, leaving three-dimensional geometric and conformational integration comparatively underexplored.
- What they did: RNAtomy jointly reconstructs pairwise geometry and residue-level descriptors using a hybrid Mamba-Transformer backbone with shifted-window attention and axis-aware feature modulation.
- The result: Across seven structural and functional tasks, its representations transfer effectively and outperform RNA foundation baselines, despite being pretrained on a substantially smaller training set.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.