TargetGAN: A generative AI framework for the design of plant core promoters with targeted activity

Plant core promoters (PCPs) are key genetic elements that control gene expression and have significant value for crop breeding and plant synthetic biology. Natural promoters (NPs) are constrained by their limited diversity and narrow activity range, and it remains unclear whether synthetic promoters (SPs) can transcend these natural constraints. Here, we present TargetGAN, a deep-learning framework trained on 76,851 NPs that integrates a generative adversarial network (GAN) with a pre-trained activity predictor to enable the de novo design of PCPs with user-defined activity. We used TargetGAN to generate 55,296 SPs and selected 5,250 for high-throughput functional validation using STARR-seq. Of these, 2,909 were successfully characterized, with a moderate correlation (Pearson correlation coefficient = 0.6435) between predicted and experimental activity. Surprisingly, 29 SPs exhibited ultra-high activity, exceeding the maximum activity of the tested NPs. Further orthogonal validation using luciferase reporter assays showed a strong positive correlation with STARR-seq measurements across a broad dynamic range. Notably, the most active synthetic candidate, SP1482, significantly outperformed the strongest tested NP, the UBI core promoter, achieving a 128-fold increase in expression relative to the 35S minimal promoter. Interpretable motif analysis suggested that ultra-high-activity promoter design can be achieved through the precise arrangement of strong activating motifs. These results demonstrate that TargetGAN is a robust and generalizable framework for the targeted generation of PCPs tailored to user-defined activity levels and will be a powerful tool both for precise gene regulation in plant systems and for overexpression analysis in genetic engineering and synthetic biology.