Designing synthetic regulatory elements using the generative AI framework DNA-Diffusion

Systematically designing regulatory elements for precise gene expression control remains a central challenge in genomics and synthetic biology. Here we introduce DNA-Diffusion, a generative artificial intelligence framework that uses machine learning trained on DNA accessibility data from diverse cell lines to design compact regulatory elements with cell-type-specific activity. We show that DNA-Diffusion generates 200-base-pair synthetic elements that recapitulate endogenous transcription factor binding grammar while exhibiting enhanced cell-type specificity. We validated these elements using a 5,850-element STARR-seq library across three cell lines. Moreover, we demonstrated successful endogenous gene modulation using EXTRA-seq, reactivating AXIN2, a leukemia-protective gene, in its native genomic context. Our approach outperforms existing computational methods in balancing functional activity with cell-type specificity while maintaining sequence diversity. This work establishes DNA-Diffusion as a powerful tool for engineering compact, highly specific regulatory elements crucial for advancing gene therapies and understanding gene regulation.