pinellolab/DNA-Diffusion
🧬 Generative modeling of regulatory DNA sequences with diffusion probabilistic models 💨
What it solves
DNA-Diffusion provides a way to generate synthetic regulatory DNA sequences that are specific to certain cell types. This allows researchers to create artificial DNA elements that mimic the properties of real regulatory regions in specific biological contexts.
How it works
The project uses diffusion probabilistic models to learn the distribution of regulatory DNA sequences. It is trained on datasets of chromatin accessible regions (such as the DHS Index dataset) and can be conditioned on discrete cell-type labels to generate new, 200bp synthetic sequences.
Who it’s for
This tool is designed for computational biologists and genomic researchers who need synthetic DNA sequences for testing or modeling regulatory elements.
Highlights
- Cell-type specific generation: Can generate sequences tailored to specific cell types (e.g., K562, HepG2).
- Customizable data: Supports training on user-provided tab-separated DNA sequence datasets.
- Flexible sampling: Includes a guidance scale parameter to tune the generation process.
- Pre-trained models: Provides model checkpoints via Hugging Face for immediate sequence generation.