The paper introduces Orthrus, a specialized RNA foundation model designed to predict the properties and functions of mature transcripts. Unlike traditional genomic models that rely on text-based learning, Orthrus utilizes a biologically inspired contrastive learning objective that leverages data from alternative splicing and evolutionary orthologs across hundreds of species. This approach allows the model to identify functional conservation and sequence importance with high efficiency, utilizing significantly fewer parameters than competing models. Orthrus achieves state-of-the-art results in predicting mRNA half-life, ribosome load, and cellular localization, even when training data is extremely limited. Furthermore, the model demonstrates a sophisticated ability to distinguish between the divergent biological roles of different isoforms originating from the same gene. By structuring its latent representation around evolutionary and functional similarities, Orthrus offers a powerful framework for advancing transcriptome research and therapeutic development.

References:

• Fradkin P, Shi R I, Dalal T, et al. Orthrus: toward evolutionary and functional RNA foundation models[J]. Nature Methods, 2026: 1-11.