README copy.md

Regulatory DNA Sequence Design

This codebase is developed on top of MDLM (Sahoo et.al, 2023).

Environment Installation

conda create -n sedd python=3.9.18
conda activate sedd
bash env.sh

# install grelu
git clone https://github.com/Genentech/gReLU.git
cd gReLU
pip install .

Data and Model Weights

All data and model weights can be downloaded from this link:

https://www.dropbox.com/scl/fi/zi6egfppp0o78gr0tmbb1/DRAKES_data.zip?rlkey=yf7w0pm64tlypwsewqc01wmfq&st=xe8dzn8k&dl=0

Save the downloaded file in BASE_PATH.

Gosai Dataset

The enhancer dataset used for this experiment is provided in BASE_PATH/mdlm/gosai_data.

Pretrained Generative Model

python main_gosai.py

The pretrained model weights are provided in BASE_PATH/mdlm/outputs_gosai/pretrained.ckpt.

Reward Oracle

python train_oracle.py

The oracle for fine-tuning is provided in BASE_PATH/mdlm/outputs_gosai/lightning_logs/reward_oracle_ft.ckpt; the oracle for evaluation is provided in BASE_PATH/mdlm/outputs_gosai/lightning_logs/reward_oracle_eval.ckpt.

The oracle for binary classification on chromatin accessibility (ATAC-Acc) is provided in BASE_PATH/mdlm/gosai_data/binary_atac_cell_lines.ckpt.

Fine-Tune to Optimize Enhancer Activity

python finetune_reward_bp.py --name test

The fine-tuned model weights are provided in BASE_PATH/mdlm/reward_bp_results_final/finetuned.ckpt

Evaluation

See eval.ipynb

Adaptations

Change the base_path in dataloader_gosai.py, finetune_reward_bp.py, oracle.py, train_oracle.py, eval.ipynb to BASE_PATH for saving data and models.

Acknowledgement

• The original dataset is provided by Gosai et al., 2023.
• The trained oracle is based on gReLU.