Abstract
End-to-end, machine-learning based design of mRNA molecules offers a powerful means to tailor their properties for specific tasks. mRNA expression level, immunogenicity, tissue specificity, stability, and localization, among others strongly depend on sequence, thus providing a rich set of properties amenable to optimization. Despite this potential, the various components of mRNA are governed by distinct grammatical and functional rules that hinder the development of a unified algorithmic approach for complete mRNA design. While machine learning and generative AI techniques demonstrate substantial benefits for individual sequence design tasks, adapting these tools to new domains or out-of-distribution tasks remains challenging. In this work, we describe a simple and powerful alteration to integrated gradients (Design by Integrated Gradients- DIGs) that serves as the foundation for several mRNA design tasks. Using this technique, we demonstrate complete model-informed mRNA design and reveal the underexplored rules governing the assembly of mRNA components into complete, high-performance transcripts. This framework closes the loop on mRNA design by unifying the strengths of predictive neural networks with the vast quantities of diverse RNA data that has been accumulating for decades across transcriptomic profiling, CLIP-seq, and individual experimental findings. By linking the benefits of neural network-based design with diverse alternative datasets, we demonstrate the design of complete mRNA sequences in out-of-distribution settings that exhibit dramatically improved translational capacity, enhanced cell-type-specific expression, and improved stability.