Multi-Objective-Guided Generative Design of mRNA with Therapeutic Properties

Therapeutic mRNA design requires simultaneous optimization of multiple competing properties including stability, translation efficiency, and expression levels—a challenge that current single-objective approaches struggle to address systematically. We introduce mRNA generation via Unrolled Trajectories and Informed Latent UpdateS (mRNAutilus), the first multi-objective guided generative model for de novo design of mRNA untranslated region (UTR) sequences. mRNAutilus combines a Masked Diffusion Model (MDM) trained on 6 million evolutionary-scale mRNA sequences with Monte Carlo Tree Search to guide generation toward Pareto-optimal solutions across multiple therapeutic properties. As objective functions, we develop XGBoost classifiers trained on mRNAutilus embeddings for predicting half-life, ribosome profiling, translation rate, and expression efficacy. Using these predictors, mRNAutilus successfully generates diverse, high-quality mRNA sequences with superior multi-objective performance compared to natural sequences while preserving coding regions for therapeutic applications, establishing a new paradigm for rational mRNA design.