DAPE-PD: Diffusion adaptation of genomic foundation model enables tailored-generation of cis-regulatory DNA

DAPE-PD adapts a pretrained Evo genomic foundation model into a bidirectional masked diffusion generator, producing functional de novo promoters validated experimentally.

• Why it matters: Programmable cis-regulatory DNA design is limited because regulatory function depends on combinatorial, distributed sequence interactions that require bidirectional context during generation.
• What they did: DAPE-PD uses pretrained genomic representations with iterative unmasking, then applies genus-specific fine-tuning on genome-derived intergenic sequences to generate regulatory DNA.
• The result: It retains genus-associated compositional features despite low similarity to fine-tuning datasets, and experimental validation identifies functional de novo promoters; in silico prioritization recovers active designs from small candidate sets across three non-model gut commensals.