Modeling ATAC data quantitatively

PoissonVI is now part of the core scvi-tools repository: https://github.com/scverse/scvi-tools.

You can call it using scvi.external.POISSONVI. Please also check out the tutorial.

Study to compare the modeling of the ATAC count data instead of the binarized matrix. All models have been implemented based on the scvi-tools suite.

To setup the environment to reproduce the results, run:

pip install -e .

• poisson_atac/: contains the code for the models, the dataloading, and plotting functions.
• poisson_atac/experiments: Experiment files to reproduce the results of the paper. Each folder contains a .yaml file with the seml configuration and a .py file with the code to run the experiment.
• notebooks: Includes an example analysis notebook for using the Poisson VAE model scvi_tools_example_notebook.ipynb and notebooks to reproduce the Figures of the paper. Additionally, we showcase the use together with Signac in scvi_tools_example_notebook_in_R.ipynb. Please follow the tutorial of scvi-tools to setup the environment.

All experiments where run through seml. The entry function is ExperimentWrapper.__init__ in the respective experiment runners experiment_runner.py.

As you will not be able to connect to the mongoDB via SEML, you have to use the provided part of the database in the respective experiments folder. To align with the notebooks, simply define your own load_config function similar to this:

import json 
import pandas as pd

def load_config(seml_collection, model_hash):
    file_path = f'{seml_collection}.json' # Provide path to json

    with open(file_path) as f:
        file_data = json.load(f)
    
    config = pd.json_normalize(file_data, sep='.')
    return config