CodonMamba: a foundation model for programmable mRNA coding sequence design

Inside every living cell, a complex assembly line turns genetic instructions into the proteins that keep life moving. These instructions are written in a language made of four chemical letters, grouped into triplets called codons. While the genetic code is universal, the way different organisms use these triplets varies; some prefer certain combinations over others, much like how different dialects favor specific words for the same idea. Scientists have long sought to rewrite these genetic messages to create powerful medicines, particularly messenger RNA, or mRNA, which acts as a temporary blueprint for making therapeutic proteins. However, designing these sequences is a delicate balancing act. The code must not only produce the correct protein but also fold correctly, remain stable, and be read efficiently by the specific host cell it enters. For years, researchers have relied on computer programs to optimize these sequences, but these tools often required a complete overhaul whenever a scientist wanted to change the rules for a new host or a different medical application.

A team of researchers has now introduced a new approach called CodonMamba, a system designed to understand and generate these genetic sequences with unprecedented flexibility. Rather than simply memorizing patterns from existing data, this system functions as a foundation model, a type of artificial intelligence trained on a vast library of natural coding sequences to grasp the underlying rules of biological language. The researchers tested this model against a comprehensive set of twelve different tasks related to predicting how mRNA behaves. In these tests, CodonMamba performed better than any previous method, ranking first in ten of the twelve categories. This success suggests the model has learned a deep, generalizable understanding of how genetic sequences function, allowing it to predict outcomes with a high degree of accuracy across a wide range of biological scenarios.

What sets this work apart is not just its accuracy, but its ability to be steered by the user at the moment of creation. Traditionally, if a scientist wanted to design a sequence for a specific organism, they had to retrain the entire computer model with new data, a slow and rigid process. CodonMamba changes this dynamic by allowing scientists to specify their preferences directly during the generation phase. By introducing user-defined rules about how often certain codons should appear, the system can instantly adapt to the needs of a specific host or application without needing to be retrained. This means a single model can be used to design sequences for different environments simply by adjusting the instructions given to it, preserving the core biological logic while shifting the stylistic choices to fit the new context.

In their experiments, the researchers demonstrated that this system could coordinate multiple design goals at once, optimizing several properties of the sequence simultaneously. They showed that the model could take a genetic design intended for one host and retarget it for another, switching the underlying preferences while keeping the essential structural choices intact. This capability marks a shift from static optimization to a programmable framework, where the design process becomes a conversation between the scientist and the model. The results indicate that this foundation model offers a precise and adaptable tool for engineering mRNA, moving the field toward a future where genetic sequences can be tailored with a level of specificity and ease that was previously out of reach.