Key points are not available for this paper at this time.
BACKGROUND: The design of mRNA drugs involves a complex and high-dimensional optimization of sequence elements to balance stability, translation efficiency, and immunogenicity. Traditional methods, often relying on local rules or limited biological templates, struggle to address the global structural dependencies within full-length mRNA. Artificial Intelligence (AI) provides a transformative approach to decode complex sequence-function relationships and navigate the vast combinatorial space for precise mRNA engineering. AIM OF REVIEW: This review systematically outlines the data infrastructure, evaluation metrics, and algorithmic frameworks driving AI-aided mRNA design. It categorizes methodologies into representation learning and generative design, specifically focusing on their applications in untranslated region (UTR) engineering, codon sequences (CDS) optimization, and the emerging trend of coordinated UTR-CDS design and integrated scoring frameworks. KEY SCIENTIFIC CONCEPTS OF REVIEW: The review highlights the evolution from discriminative models, which predict pharmacologic properties from synthetic library data, to generative frameworks (including Large Language Models) that enable de novo sequence design. It emphasizes the critical shift from isolated optimization of sub-regions to coordinated design strategies that account for cross-regional dependencies, while noting that most current methods remain sequential, modular, or scoring-based rather than fully end-to-end global models. Furthermore, the review addresses major bottlenecks such as the generalization gap between synthetic and endogenous contexts and the interpretability of "black box" models. It concludes by proposing future directions towards multimodal foundation models and multi-objective optimization to bridge the gap between computational design and clinical translatability.
Shi et al. (Mon,) studied this question.