Code in this repository compare the performace of three differentially private mechanisms: the Laplace mechanism, the exponential mechanism, and the method proposed by Johnson & Shmatikov (2013). The data used for demonstration are simulated genome-wide association study data generated by HAP-SAMPLE.

Code in this repository were used to produce results in the following paper:

"Scalable privacy-preserving data sharing methodology for genome-wide association studies." Yu, F, S. Fienberg, S. Slavković, C. Uhler (2014). Journal of Biomedical Informatics (forthcoming).

The paper is available at here and here.

The example case/control genotype data (example/case_genotypes.dat and example/case_genotypes.dat) were generated by HAP-SAMPLE with the following options:

• Population: CEU
• Source for SNPs: Chrom9_Chrom13_snps.txt
• Disease Model File: AR_chrom9_chrom13_Nov09_additive1_MAF025.txt
• Simulation Type: Case/Control
• Number of Cases: 1000
• Number of Controls: 1000
• Average breaks per cM: 1
• Output Format: SNPs v. Individuals

The disease model file describes a disease with 2 causative SNPs having addive effects. For more details about the disease model, see Malaspinas & Uhler (2010).

See http://nbviewer.ipython.org/github/fy/compare_dp_mechanisms/blob/master/notebooks/compare_all_dp_mechanisms.ipynb

Or if you have ipython notebook installed, run in the terminal

cd notebooks
../start_ipynb.sh

and open the compare_all_dp_mechanisms notebook in the browser. By default, the iPyhon notebook is at http://localhost:8888/

First of all, untar the example data:

tar xvfz exmaple.tar.gz

Then convert raw data to genotype tables:

python ./notebooks/raw_to_geno_table.py ./example/case_genotypes.dat ./example/anticase_genotypes.dat ./table.tmp

To get results from the Laplace mechanism, run:

python ./notebooks/write_chisquare.py ./table.tmp ./chisquare.tmp
python ./notebooks/get_laplace_results.py k e n_case n_control ./chisquare.tmp

where "k" is the number of top SNPs to release, "e" is the privacy budget (commonly known as epsilon in epsilon-differential privacy), and "n_case" and "n_control" are the numbers of cases and controls.

Similarly, to get results from the Exponential mechanism, run:

python ./notebooks/write_chisquare.py ./table.tmp ./chisquare.tmp
python ./notebooks/get_expo_results.py k e n_case n_control ./chisquare.tmp

To get results from the Johnson & Shmatikov method, run:

python ./notebooks/write_JS_distance.py -p pval ./table.tmp ./js_distance.tmp
python ./notebooks/get_JS_results.py k e ./js_distance.tmp

where "pval" is the p-value specified in Johnson & Shmatikov (2013), which can be interpreted as the overall p-value (say 0.05) of the multiple testing problem involving thousands of SNPs.