ARCADE: Controllable Codon Design from Foundation Models via Activation Engineering.

The synthesis of functional Messenger Ribonucleic Acid (mRNA) sequences requires the precise arrangement of nucleotide triplets to ensure efficient protein expression and long-term stability within the complex cellular environment. Prior research has shown that codon selection significantly influences the translation rate and the proper folding of the resulting polypeptide chain by affecting transfer RNA (tRNA) availability and ribosomal movement. Traditional optimization strategies often rely on fixed algorithms that target specific metrics like the codon adaptation index or sequence-wide guanine-cytosine levels to improve protein yield. These established techniques frequently struggle to balance multiple competing functional properties simultaneously during the design phase, often leading to sub-optimal sequences for sophisticated therapeutic applications. Existing computational frameworks often require extensive retraining or fine-tuning when new design objectives are introduced to the pipeline, which consumes significant time and computational resources. This absence of evidence motivated the development of a more adaptable system capable of steering sequence generation toward diverse biological targets without the need for exhaustive model updates.

This research introduces ARCADE to provide a flexible and controllable framework for generating optimized codon sequences from pretrained genomic foundation models that have learned biological syntax. The investigators sought to overcome the rigidity of current codon optimization methods by leveraging the internal representations of large-scale neural networks that already understand complex genomic patterns. The study focuses on enabling the modulation of continuous biological metrics, such as thermodynamic stability and sequence composition, without the need for model retraining or architectural modifications. By implementing activation engineering, the team aimed to manipulate the model's latent space to achieve specific functional outcomes in the output mRNA that are critical for therapeutic efficacy. The project specifically targets the improvement of mRNA vaccines and gene editing therapies through enhanced sequence programmability and the ability to meet diverse design constraints. The researchers intended to show that semantic steering vectors could effectively guide the generation process toward desired phenotypic traits by shifting the model's internal activations. This approach provides a versatile tool for synthetic biologists who require precise control over the genetic instructions they engineer for medical use.

The researchers developed the ARCADE framework by applying activation engineering techniques to pretrained genomic foundation models that capture the underlying syntax of biological sequences across various organisms. The team defined biologically meaningful semantic steering vectors within the activation space of the neural network to represent specific functional directions for sequence optimization. These vectors were designed to modulate continuous-valued properties, including the Codon Adaptation Index (CAI), which measures the usage of preferred codons, and the Minimum Free Energy (MFE). The experimental setup also included the manipulation of GC content to assess the degree of control over sequence composition and its impact on mRNA half-life and stability. The methodology avoids the computationally expensive process of retraining the foundation model for each specific design task by directly intervening in the model's hidden layers during inference. The performance of ARCADE was evaluated by comparing its output against existing codon optimization approaches across multiple design objectives to ensure its utility in real-world biological engineering. This comparative analysis allowed the team to quantify the improvements in flexibility and precision offered by their novel activation-based steering method.

ARCADE showed superior performance and significantly greater flexibility compared to traditional codon optimization methodologies across all tested biological metrics and sequence generation tasks. The implementation of semantic steering vectors allowed for the direct and precise modulation of the Codon Adaptation Index (CAI) within the generated sequences, facilitating optimal translation. The framework successfully controlled the Minimum Free Energy (MFE) of the mRNA, which is a fundamental factor for secondary structure stability and resistance to enzymatic degradation. The researchers observed that GC content could be precisely adjusted through the activation engineering process without degrading the overall sequence quality or functional potential of the mRNA. Experimental data confirmed that the foundation model's inherent knowledge could be effectively harnessed for programmable biological sequence design through simple vector-based interventions in the latent space. The results indicated that the proposed approach maintains high functional integrity while providing a broader range of controllable parameters than previous tools used in synthetic biology. These findings highlight the efficiency of using activation engineering to steer foundation models toward specific biological goals without the need for additional training data.

The findings suggest that activation engineering represents a powerful paradigm for the precise design of therapeutic mRNA sequences with tailored functional properties for clinical use. The researchers conclude that ARCADE offers a scalable solution for developing novel mRNA vaccines with optimized translation and stability profiles that can be adapted to specific viral targets. The ability to control multiple biological metrics simultaneously may accelerate the creation of more effective gene editing therapies by ensuring high expression of editing enzymes in target cells. Future applications of this technology could extend to other areas of synthetic biology where sequence-to-function mapping is essential for the design of synthetic genes and regulatory elements. The study establishes a foundation for using genomic foundation models as highly adaptable tools for programmable molecular engineering without the need for task-specific fine-tuning or retraining. The authors propose that this framework will reduce the computational burden associated with designing complex biological systems by providing a more efficient path to sequence optimization. This technological advancement could significantly shorten the development timelines for new genetic medicines and improve the precision of synthetic biology interventions.