scPortrait OT

This code implements the translation from CODEX features to single-cell gene expression. The repository contains the code for the OT-based Conditional Flow Matching algorithm, which facilitates feature translation and enables the decoder to transition from HARMONY embeddings to gene expression.

Install the software

Clone the repository:

git clone https://github.com/theislab/scPortrait_ot.git

Install the environment with conda:

conda env create -f environment.yml

Activate environment

conda activate scportrait_ot

Install environment

cd directory_where_you_have_your_git_repos/scPortrait_ot
pip install -e . 

Create an experiment folder and download the main data from scPortrait_manuscript.

Navigate the folder

• Model training is performed through a mix of hydra (configurations and sweeps), wandb (logging) and PyTorch Lightning (model implementation and callbacks).

• Model implementation can be found in scPortrait_ot. The configurations for training the reconstructing decoder and Flow Matching model are in configs_autoencoder and configs. The training processes are decoupled and can be run independently.

• To train models, make sure to complete all the configuration files with the paths to the data and target folder according to your file hierarchy. The files requiring path changes are:

  • configs/datamodule/tonsilitis_cite.yaml
  • configs/training_config/default.yaml
  • configs/configs_autoencoder/train_cite_log.yaml

• The model training scripts are in scripts; they automatically run a hyperparameter search on slurm. There are two subfolders, one for training the Flow Matching model and one for the decoder model. Make sure to create a sub-folder called logs inside these folders if running the scripts on slurm.