codonGPT: Reinforcement learning on a generative language model optimizes RNA sequences under biological constraints

Emerging generative models for biology focus on DNA, non-coding RNA, or proteins, ignoring information hidden in mRNA. Additionally, in protein engineering and mRNA therapeutics the design of mRNA sequences is still a challenge, lacking a clear framework. Here, we introduce and rigorously evaluate two novel methods: a foundational model for mRNA and a reinforcement learning mRNA design framework built on such a model. codonGPT is the first generative foundational language model trained directly on coding mRNA sequences. To solve the problem of synonymous constraints that are only unique to mRNA, we introduce a novel method of inference-time masking, along with house-keeping genes evaluation. For the first time, we also rigorously demonstrate, that for precise mRNA therapeutics design, reinforcement learning on such a model provides a clear framework for biological optimization. Our study introduces a novel foundational model for mRNA and a new reinforcement learning based paradigm for mRNA sequence design.