Modeling the Language of Codons with Artificial Intelligence

The messenger RNA (mRNA) sequence plays a central role in expressing functional proteins from genes. Species-specific nonuniform codon choices influence key processes, including mRNA stability, translation accuracy and efficiency, and cotranslational folding. As evolutionarily selected codon sequences encode multiple overlapping, yet often subtle, signals, advanced deep learning models are needed to identify their statistical patterns. This review summarizes recent progress in artificial intelligence (AI) models that operate at the codon level, including both discriminative and generative approaches. We discuss codon language models and their application to downstream prediction tasks, as well as generative models used to design codon sequences for heterologous expression and to recode proteins for improved mRNA stability and expression. Finally, we highlight studies leveraging codon-based AI models to gain insights into molecular evolution and outline several open problems in this rapidly developing field.