/synthetic-interventions

Primary LanguageJupyter Notebook

Causal Imputation via Synthetic Interventions

This code accompanies Causal Imputation via Synthetic Interventions. See here for the citation.

The files are organized as follows:

  • src contains the baseline algorithms and the synthetic interventions algorithm
  • processing contains scripts for processing the raw data
  • evaluation contains classes for evaluating the performance of various algorithms
  • scratch contains temporary files for one-off tasks, such as checking that new code works

Replicating data processing from the raw data [Optional]

The data files are too large (5Gb) to be kept in the repository, but can be downloaded by running download.sh.

Once you have downloaded the data, you can run the processing scripts in the correct order via

bash process.sh

This will create ~44Gb of processed data, and may take about 1 hour.

Replicating algorithm results and plots

For convenience, the processed data is available at the following anonymized link: https://drive.google.com/drive/folders/1WKZRHY-v2zgu6XZqXLiCwEArqP4KfziR?usp=sharing. Copy the contents of this folder to data/processed/

To re-create the results from the paper, first set up a virtual environment with the necessary packages by running

bash setup.sh

Then, run the algorithms and the plotting code via

source venv/bin/activate
bash create_plots.sh

Because of the reliance on parts of the raw dataset, this script does not include calls to the functions which create the UMAP plots and the plots of which cell/drug pairs are available. This will be remedied when publicly releasing the code.