Designing DNA With Tunable Regulatory Activity Using Discrete Diffusion.

Area of Science:

• Genomics
• Synthetic Biology
• Computational Biology

Background:

• Designing regulatory DNA with precise control is crucial for biotechnology and medicine.
• Deep generative models show potential for sequence design, but their ability to replicate natural regulatory DNA's specificity is unclear.

Purpose of the Study:

• Introduce DNA Discrete Diffusion (D3), a novel generative model for designing regulatory DNA.
• Evaluate D3's performance in generating sequences with target activity, distribution, and composition.
• Assess D3's utility in scenarios with limited training data and its ability to learn predictive sequence representations.

Main Methods:

• Developed D3, a generative model employing an iterative nucleotide-substitution process.
• Benchmarked D3 against diffusion baselines using computational metrics.
• Validated D3-designed sequences in K562 cells using lentiviral promoter-reporter assays (lentiMPRA).
• Investigated D3's performance with limited data and its capacity for learning sequence representations.

Main Results:

• D3 demonstrated superior performance in regulatory sequence generation compared to diffusion baselines.
• D3-designed sequences showed measurable regulatory activity and better recapitulated natural activity distributions in experimental assays.
• D3 effectively generated informative sequences even with scarce labeled data, improving predictive models.
• When trained without activity labels, D3 learned sequence representations predictive of enhancer activity, comparable to existing genomic language models.

Conclusions:

• D3 provides a practical framework for designing synthetic regulatory DNA.
• The model offers insights into sequence features driving context-specific regulatory activity.
• D3 enhances the design and understanding of functional genomic elements.