This is the official repository of MorphoDiff, a diffusion based generative pipeline to predict high-resolution cell morphological responses under different conditions based on perturbation encoding. MorphoDiff is capable of producing guided, high-resolution predictions of cell morphology that generalize across both chemical and genetic interventions.
The requirements.txt file provided enables installation of packages used for training MorphoDiff model:
- create a virtual environment:
conda create -n morphodiff python=3.10 -y - install dependences:
pip install -r requirements.txt
The scripts/train.sh provides commands for defining parameters required for training MorphoDiff and Stable Diffusion, with description of each parameter provided in the bash script. After defining the parameters of the training script, run the following for submitting the training job using Slurm.
sbatch scripts/train.shYou should provide a path to a pretrained Stable Diffusion model to the script. It can either be the pretrained Stable Diffusion (such as stable-diffusion-v1-4 originally used for training MorphoDiff), or one of the pretrained MorphoDiff checkpoints (link will be provided soon).
After the training is completed for the specified number of steps, the scripts/train.sh automatically resubmits the job and resumes training from the last checkpoint. You can set the total_steps parameter to not train more than a specific number of steps, or comment the scontrol requeue $SLURM_JOB_ID line that resubmits the job once it is finished.
Data folder contents must follow the structure described in https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a metadata.jsonl file must exist to provide the perturbation id for the images. The perturbation ids used in MorphoDiff analysis are provided in the required_file/ folder.
The raw datasets used in the MorphoDiff paper can be downloaded from the following websites.
| Dataset | Link |
|---|---|
| RxRx1 | https://www.rxrx.ai/rxrx1 |
| BBBC021 | https://bbbc.broadinstitute.org/BBBC021 |
| Rohban et al. | https://github.com/broadinstitute/cellpainting-gallery - (cpg0017-rohban-pathways) |
For each dataset, there should be a separate csv file in the required_file/ directory that contains the perturbation id (in the first column, perturbation id must be in the same format as what is provided in the metadata.jsonl), and perturbation embedding with numerical values of the pertubration encoding provided in the following columns. Please look at sample csv files provided.
We used scGPT single-cell foundation model for encoding genetic perturbations (follow installation instruction from https://github.com/bowang-lab/scGPT), and RDKit tool (https://github.com/rdkit/rdkit) for encoding chemical compounds. The code for converting gene ids and SMILES representation of compounds to pertubration embedding is provided in the preprocessing/ folder.
The scripts/generate_img.sh script is a Slurm based bash script that takes the path to the pretrained checkpoint, a file of perturbation list, the number of images to generate per perturbation, and an address to save the images of each perturbation in a separate folder. You should set the parameters described and documented in the scripts/generate_img.sh and run it as follow
sh scripts/generate_img.shThe CellProfiler pipeline used for extracting CellProfiler features is provided in cellprofiler/pipeline/ folder. All scripts for CellProfiler feature preprocessing, and validation are provided in the cellprofiler/ folder.