Data Generation

Code to generate the data included in [FILE PATH].

from sc_causal.causal.util import create_in_distribution_h5ad, create_out_of_distribution_h5ad

fname_iid = 'out_iid_dataset' # Name of output directory
create_in_distribution_h5ad(fname_iid)

fname_ood = 'out_ood_dataset'
create_out_of_distribution_h5ad(fname_ood)

Running Baselines

To run baselines, cd to the sc_causal/GEARS/demo directory and run

python run_baselines.py --id ood_split_id

where ood_split_id is a number, either [0, 1, 2, 3, 4, -1]. Setting ood_split_id = -1 runs the in-distribution task.

Training

To train SCCVAE:

cd sc_causal
python causal/run.py -x 'out_filename' -d 0 --ood --split-num 0

Flags:

• x Name of output directory.
• -d Which CUDA device to run the code.
• --ood If this flag is included, runs the OOD split given by --split-num. Otherwise, runs the IID task.
• -m Changes the graph in the SCM. Can be full, causal, conditional, or random.
• --random-graph-seed If training a random graph, keeps the generated graph consistent across all OOD splits.

Evaluation

SCCVAE evaluation consists of two steps: Shift selection (for OOD tasks) and inference.

Shift Select (OOD only)

In sc_causal/causal/select_model_shift_values.py, edit the list SAVEDIRS to be the list of all models to evaluate shift selection for, specifying the OOD split number, the model directory, and the sparse graph. Then, while in the sc_causal directory, run

python causal/select_model_shift_values.py

Inference

python causal/inference.py -m 'out_filename' -s 'descrption' -n 'load_model_name' -d 0 --data-split 'test'  --ood-split 4

Flags

• m Model output directory.
• s Descriptive file name for saving output adata.
• n Model name to load from. E.g., best_val_mmd_shiftselect_hard
• d CUDA device.
• --data-split Evaluating on train, test, or val datasets.
• --ood-split If OOD, which OOD split is being evaluated.

Figure replication

To reproduce the figures from the paper.

First, download zip files from https://drive.google.com/drive/folders/1rxHl_4EPdoLA-ahyO7Woa9kTa4UOhUrj?usp=sharing. h5ad_datafiles.zip and outs.zip go in sc_causal, and all .zip files under the subfolder GEARS files go under sc_causal/GEARS/demo.

Starting in the root directory,

cd sc_causal/visualize_paper_tables_and_figures
python tables.py
python fig2a.py
python fig2b.py
python fig3_fig7.py
python fig4_fig8.py
python fig5a.py
python fig5b.py
python fig6.py
python fig9.py