Orthrus: toward evolutionary and functional RNA foundation models.

Orthrus, a Mamba-based mature RNA foundation model, learns latent representations that cluster RNA sequences by functional and evolutionary similarity using splice-isoform and orthologous-gene transcript pairs.

• Why it matters: Predicting mature RNA properties that drive transcript function and regulation is still limited despite rapidly growing genomic data, and existing RNA foundation models miss biological domain knowledge.
• What they did: Orthrus is pretrained with self-supervised contrastive learning using biological augmentations, maximizing embedding similarity between transcripts from splice isoforms of ten model organisms and orthologous genes across 400+ mammalian species.
• The result: Its mature RNA isoform representations outperform genomic foundation models on mRNA property prediction tasks and capture divergent biological function of individual transcript isoforms, using only a fraction of fine-tuning data.

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