Directed optimization and generation of yeast promoter sequences driven by deep learning

Promoter design and optimization are critical for precise control of gene expression. However, existing methods remain constrained by insufficient understanding of sequence features and a dependence on pre-trained sequence-to-expression (S2E) predictors. To address this limitation, we present DOSDiff, a diffusion-based framework that learns promoter features and enables targeted optimization through local sequence editing without requiring an S2E model. DOSDiff controllably generates desired promoters by capturing promoter encoding rules, achieving a 4-mers distribution similarity of 0.8910 ± 0.0002 (n = 3), surpassing WGAN-GP (0.7282 ± 0.1103) and DBGM (0.7549 ± 0.0274) by 22.36 % and 18.03 %, respectively. In Saccharomyces cerevisiae, locally optimized genes of interest achieved one-to-one expression enhancement. Additionally, DOSDiff demonstrated generalization across different yeast species, with all optimized Pichia pastoris promoters retaining functional activity and achieving up to 1.70-fold expression gain in vivo validation. These results highlight DOSDiff as a promising and adaptable framework for precise promoter engineering.