DAGGER

Code for replicating the experiments in the paper DAGGER: A sequential algorithm for FDR control on DAGs by Aaditya Ramdas, Jianbo Chen, Martin J. Wainwright, Michael I. Jordan.

Dependencies

The code for DAGGER runs with Python 2.7. Please pip install the following packages:

numpy
scipy
statsmodels
matplotlib
networkx

Or you may run the following and in shell to install the required packages:

git clone https://github.com/Jianbo-Lab/DAGGER
cd DAGGER
sudo pip install -r requirements.txt

To run the experiments in Section 4.1 of the paper, R is required. Please install the following packages in R.

igraph
cherry
globaltest in Bioconductor

Running experiments in Section 4.1

We provide codes for reproducing figures in Section 4.1. Run the following commands in shell:

###############################################
# Omit if already git cloned.
git clone https://github.com/Jianbo-Lab/DAGGER
cd DAGGER
############################################### 
cd GO
# Generate simulated data (hypothesis and p-values on the GO graph.)
Rscript make_data.R

# Generate all the jobs. 
python generate_jobs.py

# Run an example job.
# set_name: full or cellprolif. which set of nodes on the GO graph is used.
# seed: random seed.
# pvalues_type: simes or ind: whether to use Simes' p-values or independent p-values.
# algorithm: DAGGER,lord,BH,all-goeman,any-goeman,SCR.DAG,BH.DAG,Holm 
# pi0: proportion of nulls. 
python run_simulations.py --set_name full --seed 1 --pvalues_type simes --algorithm DAGGER --pi0 0.5 
# Run all jobs.
sh jobs.sh

# Plot figures.
python plot.py

The generated plots can be found in DAGGER/graph_structure/results. See core/algorithm.py and GO/run_simulations.py for details.

Power comparison for full and subgraph GOs with independent and Simes p-values.

Running experiments in Section 4.2 and Section 4.3

We provide codes for reproducing figures in Section 4.2 and Section 4.3. Run the following commands in shell:

###############################################
# Omit if already git cloned.
git clone https://github.com/Jianbo-Lab/DAGGER
cd DAGGER
############################################### 
cd graph_structure
# Running experiments in Section 4.2.
c```

|*<center>BH vs. DAGGER. </center>*|
|:--:| 
|![alt-text-3](https://github.com/Jianbo-Lab/DAGGER/blob/master/figures/bh.png?raw=true)|

```shell
# Running experiments in Section 4.3
# Mountain vs. Valley. 
python run_simulations.py --experiment mountain_vs_valley --n_replications 10

Mountain vs. Valley.

# Running experiments in Section 4.3
# Shallow vs. Deep.  
python run_simulations.py --experiment shallow_vs_deep --n_replications 10

Shallow vs. Deep.

# Running experiments in Section 4.3
# Diamond vs. Hourglass.  
python run_simulations.py --experiment diamond_vs_hourglass --n_replications 10

Diamond vs. Hourglass.

The generated plots can be found in DAGGER/graph_structure/results. See core/algorithm.py and graph_structure/run_simulations.py for details.

Citation

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

@article{ramdas2017dagger,
  title={DAGGER: A sequential algorithm for FDR control on DAGs},
  author={Ramdas, Aaditya and Chen, Jianbo and Wainwright, Martin J and Jordan, Michael I},
  journal={arXiv preprint arXiv:1709.10250},
  year={2017}
}

Acknowledgements

We thank Jelle Goeman for helping running code from his packages and reproducing results in his papers. We also thank Lihua Lei for sharing code for SCR-DAG and BH-DAG.

Jianbo-Lab / DAGGER