Diffusion-Based Generative Network for de Novo Synthetic Promoter Design

Area of Science:

• Synthetic Biology
• Computational Biology
• Genomics

Background:

• Computer-aided design is advancing synthetic promoter engineering.
• Deep learning models are used for evaluating synthetic promoters, but de novo design remains underexplored.

Purpose of the Study:

• To explore generative models for de novo promoter design.
• To establish a diffusion-based model for Escherichia coli promoter generation.

Main Methods:

• Developed a diffusion-based generative model trained solely on sequence data.
• Improved the Fréchet Inception Distance (FID) calculation using a convolution layer for feature extraction.

Main Results:

• The model learned essential characteristics of natural promoters.
• Generated synthetic promoters exhibit structural and compositional similarity to natural ones.
• Achieved an FID of 1.37, indicating high similarity between synthetic and natural promoter distributions.

Conclusions:

• A data-driven generative model is feasible for de novo promoter design.
• This approach offers a new method for creating synthetic promoters with natural-like properties.