Deep Learning-Guided Reverse Translation Enhances Soluble Expression of Recombinant Proteins in Escherichia coli

Enhancing the soluble expression of heterologous proteins in chassis microorganisms is critical for fundamental biological research and synthetic biology-driven industrial applications. Current methods for designing DNA sequences to ensure high soluble expression often rely excessively on high-frequency codons while overlooking optimal codon context, leading to suboptimal outcomes. To address these limitations, we developed an integrated deep learning framework combining a synonymous codon generation (SCG) model and a gene expression level prediction (GELP) model. The SCG model captures codon usage patterns in Escherichia coli using large-scale genomic data, whereas the GELP model leverages gene expression data to prioritize sequences with high soluble expression potential. We validated our approach by optimizing the DNA sequences of two industrial enzymes, α-glucan phosphorylase (αGP) and isoamylase (IA), achieving significant and reproducible improvements in soluble expression (mean 12.2–16.9-fold, n = 3 and 2.6–3.4-fold, n = 4), confirmed by one-way ANOVA and one-sample t-tests. This study provides a useful tool for designing DNA sequences that confer high soluble expression and for understanding the relationship between DNA sequence and protein expression. Notably, SCG-GELP reveals a core-avoiding codon optimization strategy that substantially enhances soluble protein yield.