CyanoDiff

Official repository for "CyanoDiff: Class-Conditional Cyanobacterial Promoter Generation via Masked Diffusion Language Modeling".

CyanoDiff is a masked diffusion language model for designing cyanobacterial promoters with controllable expression strength. Pretrained on cyanobacterial genomes and fine-tuned on labeled promoters, it generates 81 bp promoters steered toward a low / mid / high expression class via classifier-free guidance.

git clone https://github.com/Passion4ever/CyanoDiff.git
cd CyanoDiff
conda create -n prom python=3.10 -y
conda activate prom
pip install torch --index-url https://download.pytorch.org/whl/cu121
pip install -e .

Pretrain on genomes:

python scripts/pretrain.py --config configs/pretrain.yaml

Fine-tune on one of the four built-in species (--species selects 7120 / 6803 / MED4 / MIT9313 and expands its data paths):

python scripts/finetune.py --config configs/finetune.yaml \
    --species 7120 --pretrain path/to/pretrained.pt

To fine-tune on your own species, prepare train.csv / val.csv / test.csv, each with two columns:

• sequence — an 81 bp promoter (A/T/G/C), 60 bp upstream + TSS + 20 bp down
• log_expression — log10(reads + 1) of that promoter's expression

Point the config at them and run without --species (expression is binned into low / mid / high quantile classes automatically from your training set):

python scripts/finetune.py --config configs/finetune.yaml \
    --pretrain path/to/pretrained.pt --run-name my_species \
    # set data.{train,val,test}_csv in configs/finetune.yaml to your CSVs

Generate sequences per expression class across guidance scales:

python scripts/generate.py --checkpoint path/to/finetuned.pt \
    --scales 1 --n-samples 1000

This writes per-class CSVs and a metrics.json (GC content, GCGATCGC motif rate, k-mer JSD, diversity, novelty, uniqueness).

Genomes, processed promoter datasets, and trained checkpoints are not in this repository — they are hosted on Zenodo:

• Zenodo: https://doi.org/10.5281/zenodo.20590300

The deposit bundles data.tar.gz (486 genomes + four-species promoter datasets and their source tables), checkpoints.tar.gz (pretrained model + four fine-tuned models), and code.tar.gz (this model-core package). Download and extract into the repository root:

tar -xzf data.tar.gz
tar -xzf checkpoints.tar.gz

If you use CyanoDiff, please cite the dataset deposit (a paper citation will be added upon publication):

@dataset{cyanodiff_2026,
  author    = {Yang, Guang and Li, Jianing and Kwoh, Chee Keong and
               Hu, Jinlu and Shi, Jian-Yu},
  title     = {{CyanoDiff: data, model checkpoints, and code for
               class-conditional cyanobacterial promoter generation via
               masked diffusion language modeling}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20590300},
  url       = {https://doi.org/10.5281/zenodo.20590300}
}

See LICENSE.