Privacy-preserving Genotype Imputation based on Homomorphic Encryption

This is a solution from the SNU team for privacy-preserving genotype imputation based on Homomorphic Encryption (HE). Basically it consists of four directories: /ModHEaaN, /data_origin, /plain and /encrypted. To run our HE-based genotype imputation solution, follow the instruction described below in order.

Download Datasets

At first, we need to download four datasets "Total", "EUR", "AMR" and "AFR". Please follow the steps below:

1. Download original data files from here and save in folder /data_origin.
2. Download modified data files from here and save in folder /plain/data_mod.

• Total: /data_origin/*.txt, /plain/data_mod/Total_mod/*.txt
• EUR: /data_origin/*_EUR.txt, /plain/data_mod/EUR_mod/*.txt
• AMR: /data_origin/*_AMR.txt, /plain/data_mod/AMR_mod/*.txt
• AFR: /data_origin/*_AFR.txt, /plain/data_mod/AFR_mod/*.txt

Build Plain Models

We generate 1-hidden layer neural network models for several datasets represented by population. For each dataset, one can choose different models determined by window_size, which denotes the number of adjacent tag SNPs for each target SNP. Experiments based on the given dataset shows that the choice window_size = 40 provides the best accuracy of genotype imputation. Note that the larger window_size implies the higher computational cost. We also provide the multi-processing option for the acceleration so that one can dynamically choose the number of processes based on his/her computer environment.

To generate the imputation models, command the following:

$ cd ./plain
$ python New_gen_model_W.py -p <population> -w <window_size> -n <number_of_processes>

• population: Total, EUR, AMR, AFR
• window_size: 8, 16, 24, 32, 40, 48, 56, 64, 72
• number_of_processes: 1, 2, 4, 8, 16

The generated plain models will be saved in the /encrypted/<population>_DNNmodels/DNNmodels_<window_size>_c directory.

Model for Low Minor Allele Frequency data

We have added the model for data with low minor allele frequency (MAF). This data consists of 117k number of target SNPs. To run with low MAF data, run New_gen_model_w_117k instead of original one and population=Total. Then, the generated plain models will be saved in the /encrypted/Total_DNNmodels/DNNmodels_lowMAF_<window_size> directory.

Build ModHEaaN

For the encryption of test data, we use the ModHEaaN library, which is a light-version implementation of the approximate HE scheme CKKS. Contrary to the original implementation of the CKKS scheme HEAAN, our ModHEaaN library does not have any dependency on multi-precision libraries GMP and NTL, and only supports homomorphic addition and 1-depth constant multiplication (hence bootstrapping disabled).

To build ModHEaaN, command the following:

$ ./ModHEaaN/heaan
$ cmake CMakeLists.txt
$ make all

Build the HE-based Genotype Imputation executable

$ ./encrypted/impute_dnn
$ cmake CMakeLists.txt
$ make all

The executable file is generated as enc_impute in the directory ./encrypted/impute_dnn.

Run enc_impute

The command line to execute enc_impute depends on the target dataset.

For Total Dataset

One need to choose two input arguments window_size and number_of_targetSNP. window_size is the number of adjacent tag SNPs for each target SNP, and number_of_targetSNP represents the number of target SNPs.

$ cd ./encrypted/impute_dnn
$ ./enc_impute <window_size> <number_of_targetSNP>

• window_size: 8, 16, 24, 32, 40, 48, 56, 64, 72
• number_of_targetSNP: 20, 40, 80, 117

For instance, The argument below runs the genotype imputation of 80k target SNPs, with window size 40.

$ ./enc_impute 40 80

When you want to deal with low MAF data, define number_of_targetSNP = 117, then the other parameters are the same as original Total data.

For EUR/AMR/AFR Dataset

In the case of EUR/AMR/AFR Dataset, the command line is fixed as

$ ./enc_impute populations 80

The output will be genotype scores on EUR, AMR, and AFR datasets for the fixed window_size=40 and number_of_targetSNP=80.

• NOTE: The running time linearly grows up in terms of window_size, but the accuracy does not. For the given datasets, the choice window_size=40 shows the best accuracy.

Measure the Accuracy (MicroAUC)

If you succeed to run our solution, then the genotype score results of our solution "genotype_score" will be saved in /encrypted/impute_dnn denoted by score_window<window_size>_<number_of_targetSNP>k.csv. Note that the real genotypes of test data "genotype_real" is saved in /plain denoted by real_<number_of_targetSNP>k.csv. To run evaluation.py in the ./plain directory, command

python3 evaluation.py -i <genotype_score> -t <genotype_real> -o output.png

e.g. After running $ ./enc_impute 40 80 in the previous step, then

$ cd ./plain
$ python3 evaluation.py -i ../encrypted/impute_dnn/score_window40_80k.csv -t real_80k.csv -o output.png

e.g. After running $ ./enc_impute populations 80 in the previous step, then

$ cd ./plain
$ python3 evaluation.py -i ../encrypted/impute_dnn/score_EUR_80k.csv -t real_EUR_80k.csv -o output_EUR.png
$ python3 evaluation.py -i ../encrypted/impute_dnn/score_AMR_80k.csv -t real_AMR_80k.csv -o output_AMR.png
$ python3 evaluation.py -i ../encrypted/impute_dnn/score_AFR_80k.csv -t real_AFR_80k.csv -o output_AFR.png