CodonRL: Multi-Objective Codon Sequence Optimization Using Demonstration-Guided Reinforcement Learning

CodonRL is a reinforcement learning framework for multi-objective mRNA codon optimization that learns a structural prior from efficient folding feedback and demonstration-guided replay. The framework uses LinearFold for fast intermediate reward computation, milestone-based rewards to address delayed feedback in long-range optimization, and enables user-controlled trade-offs between translation efficiency, RNA stability, and compositional properties. On a benchmark of 55 human proteins, CodonRL outperforms GEMORNA by 9.5% higher CAI, 25.4 kcal/mol more favorable MFE, and 3.4% lower uridine content.

Recommended citation: Du S, Kaynar G, Li J, You Z, Tang S, Kingsford C. CodonRL: Multi-Objective Codon Sequence Optimization Using Demonstration-Guided Reinforcement Learning. bioRxiv. 2026. doi:10.64898/2026.02.12.705465
Download Paper