Fast Multi-objective RNA Optimization with Autoregressive Reinforcement Learning

Codon optimization is essential in mRNA vaccine development, while existing tools face limitations in the computational efficiency, sequence diversity and universality. To address these challenges, we develop RNAJog (RNA Joint Optimization with autoregressive Generative model), a framework integrating autoregressive generation with reinforcement learning to optimize codon sequences for minimum free energy (MFE), codon adaptation index (CAI) and GC content, even enabling sequence design without requiring annotated training data. Evaluations in both in silico and wet-lab experiments have confirmed RNAJogs effectiveness and efficiency, with two orders of magnitude faster than traditional algorithm (LinearDesign) for long RNA sequence and about a 10-fold increase in antibody titer compared to the wild-type mRNA for Influenza virus hemagglutinin (HA) mRNA vaccine design in mouse. RNAJog also supports biological constraints for sequence optimization. Using this feature, we minimized m6A modification motifs in Bmp2 coding sequence for enhancing the translational efficiency and RNA stability, which are validated in cell-based experiments.