Unconditional de novo generation and functional validation of synthetic promoters with a diffusion model

Generative neural networks enable de novo design of synthetic genetic elements. Here, we present DiPro, an unconditional diffusion model that learns E. coli promoter features to generate novel sequences without conditional input. DiPro directly generates 50-bp promoter sequences from Gaussian noise using a time-conditioned Transformer denoiser, without auxiliary discriminators or recognizers. DiPro accurately recapitulates natural promoter statistics, including sequence motifs, k-mer spectra, conserved -35/-10 spacing (16-18 bp). Over 90% of generated sequences were classified as real promoters by an independent recognizer, indicating high structural fidelity. To design stationary-phase-specific promoters, generation was guided by inpainting randomly masked regions (three non-overlapping 6-mers, 18 bp) within sigma38-associated (σ³⁸) seed sequences, yielding 5000 candidates while preserving global promoter architecture. Bayesian classification resolved these sequences into three distinct clusters with divergent -35/-10 composition and motif conservation. Experimentally, 81.6% of tested promoters exhibited stationary-phase activity, with the strongest variant reaching 79.6% of the activity of the constitutive promoter J23119. Promoters displayed distinct activation timing, reflecting heterogeneous stress responses. As a proof-of-concept, selected promoters were used to drive a lysis protein, achieving phase-specific and tunable cell lysis that matched promoter activation timing, with early-phase promoters reducing culture density to an optical density at 600 nm of approximately 1.5-2.0. Together, these results establish DiPro as a generative framework for creating functional bacterial promoters with programmable temporal dynamics, offering a new tool for synthetic biology and metabolic engineering.