Deep learning has enabled predictive modeling of mRNA translation from 5′ UTRs, but current state-of-the-art models, such as Optimus 5-Prime, require hundreds of thousands of labeled sequences, limiting their accessibility and scalability. We present a data-efficient foundation model that achieves comparable predictive accuracy to Optimus 5-Prime using less than 1% of its training data. Our approach leverages parameter-efficient architectures and unsupervised sequence pretraining to model sequence-to-translation relationships with minimal supervision. In a major extension beyond translation, we demonstrate that the same model can predict mRNA stability from 3′ UTR sequence features, enabling dual-objective optimization of gene expression. This unified model provides an end-to-end predictive framework for both protein production and transcript longevity, two critical dimensions of therapeutic mRNA and synthetic biology design. Our results offer a scalable, cost-effective path toward general-purpose RNA regulatory models and suggest a paradigm shift in how functional sequence-labeling data is leveraged for biological model training.
