Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View

Paper by Haoyue Dai, Ignavier Ng, Gongxu Luo, Peter Spirtes, Petar Stojanov, Kun Zhang. Appears at ICLR 2024 (oral).

The first to deal with dropouts fully nonparametrically: conditional independence (CI) relations in the data with dropouts, after deleting the samples with zero values for conditioned variables, are identical to the CI relations in the original data.

1. To reproduce the simulation results in the paper:

1.1. Generate the random graphs:

Run the block of line 507-516 in main.py. The graphs will be saved in the folder ./graphs/.

1.2. Simulate the data:

Run the block of line 528-560 in main.py. The data are simulated according to the generated graphs in step 1, with different SEM models and different dropout mechanisms. Generated data will be saved in the folder ./data/.

1.3. Run the experiments:

Below is an examplar command to run:

python main.py --dagid 0
               --gtype ER
               --funcform linear
               --exonoise gaussian 
               --nonnegative raw 
               --dropouts latent.CAR.truncated.logistic

where different configurations to run the experiments can be specified by the arguments. The results will be saved in corresponding subfolders in ./data/.

2. To run PC or GES with zero-deletion on your own scRNA-seq data:

Simply initialize the corresponding CI tester and then pass to PC or GES. Here is an example:

tester = cit.CIT(
    your_own_data, # np.ndarray, in shape (n_samples, n_features)
    method="zerodel_fisherz" # or "zerodel_kci"
)
cg = pc(
    your_own_data,
    alpha=0.05, 
    indep_test=tester
)

For more details on visualizing / benchmarking the results, please refer to the run_pc (line 441) and run_ges (line 471) functions in main.py.

For the details of reframing the GES by CI tests without any score functions (and for sample size correction), please refer to e.g., line 521-556 block in ./csl/search/ScoreBased/GES.py (implemented based on this repo).

3. To incorporate zero-deletion into other GRNI methods:

Please refer to e.g., line 177-228 block and line 255-294 block in ./csl/utils/cit.py for the details in zero-deleted CI tests (implemented based on this repo).

Note that given our causal dropout models, the current codes only delete data samples with zeroes in conditioning variables. You may also try to delete data samples with zeroes in all involved variables (i.e., X, Y, and the conditioning variables), which is still correct (according to Def 2 and Thm 2 in the paper), though it may be less powerful with fewer sample size.

Citation

If you use this code for your research, please cite our paper:

@inproceedings{dai2024gene,
  title={Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View},
  author={Haoyue Dai and Ignavier Ng and Gongxu Luo and Petar Stojanov and Peter Spirtes and Kun Zhang},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://openreview.net/forum?id=gFR4QwK53h}
}