CodonMamba: a foundation model for programmable mRNA coding sequence design

Abstract

Although mRNA codon language models provide a generalizable framework for biological sequence design, effective CDS design requires both a learned sequence design space that captures biological constraints and context-configurable design preferences. Here we present CodonMamba, a codon language model framework for mRNA prediction and programmable CDS design. Pretrained on large-scale coding-sequence corpora, CodonMamba achieved state-of-the-art performance across a comprehensive benchmark of 12 mRNA prediction tasks, ranking first on 10 of 12 tasks. Furthermore, CodonMamba establishes a programmable CDS design framework, transforming codon optimization from a process dependent on preferences embedded during model training into an inference-time steerable generation framework. By introducing user-specified codon usage priors during generation, CodonMamba enables inference-time steering toward host- or application-specific codon preferences without retraining. In design experiments, CodonMamba enabled coordinated optimization over multiple design-relevant sequence properties and demonstrated programmable cross-host CDS retargeting through inference-time prior switching, while preserving most model-derived codon choices. Together, these results establish CodonMamba as a programmable foundation model framework for CDS design, enabling context-specific mRNA sequence engineering and highlighting a promising direction for precise mRNA design using foundation models.