Predictor-Guided Latent Space Codon Optimization for Maximizing Protein Expression
This study addresses the challenges of navigating the vast discrete combinatorial space in codon optimization and the limited predictive accuracy of traditional heuristic methods for protein expression. We propose LSCO, a framework that maps discrete sequences into the latent space of a pretrained mRNA language model to enable continuous optimization via gradient-based search. Furthermore, LSCO incorporates an uncertainty-aware predictor, minimum free energy regularization, and a naturalness prior derived from protein back-translation, effectively overcoming conventional limitations. Wet-lab validation on antibody datasets demonstrates that LSCO significantly outperforms frequency-based baselines and deep generative models in expression level prediction while preserving favorable biophysical properties.