Homomorphically Encrypted Outsourced Genotype Imputation with Lattigo

This repository hosts the code for the solution submitted by the EPFL-LDS team (https://lds.epfl.ch) to iDash 2019 track II (https://humangenomeprivacy.org/2019), evolved and improved for the article "Secure and Practical Genotype Imputation Methods via Homomorphic Encryption". The code implements a secure genotype imputation service built on top of the Lattigo lattice-based homomorphic encryption library (https://lattigo.epfl.ch), developed by LDS.

Instructions follow below for building the docker and reproducing the timings and performancd figures shown in the aforementioned paper.

How to build the Docker

docker build -t epfl_solution .
docker run -it epfl_solution

Running performance evaluation

In order to obtain the execution times of the system for prediction in the test set, you can run

make perf 2>&1 | tee timings.txt

Running accuracy evaluation

In order to obtain the AUC (Area Under Curve) figures for the predictions in the test set, you can run

make light 2>&1 | tee aucLight.txt

By default, the script obtains the result for regressions of 64 coefficients (63 neighboring SNPs + 1 intercept variable), and this can be changed in the script to report the accuracy for different regression sizes.

Data files

There files included in the repository are structured in different folders, depending on whether they are used for re-training the used models, or for testing their performance and accuracy:

Model training

Under the iDash 2019 Track II scenario, the model training happens offline, in a trusted environment. The python scripts required to rerun the training are in the training folder. In order to benchmark the sytem with the available training-test dataset, re-training is not needed.

Prediction/Imputation (test)

• The plaintext models (regression coefficients) trained on the sample data are in prediction/prediction_data

• All the original files for testing (including the population) are in prediction/prediction_data/original_data

• Plaintext models for population stratification are chosen between [trained for all] matrices and [trained for population x] matrices according to the highest prediction results

Result files

Results are saved in the prediction/results subfolder.

• When running the system for imputation, the direct results are saved in binary format in ypred.binary (for storage efficiency consideration).

• After running make auc, the prediction results are saved in the folder prediction/results, with format ypred_window[4/8/16/32/48/64].csv, and population stratification results are files with name ypred_[AFR/AMR/EUR]_window[4/8/16/32/64].csv

• All the mentioned non-binary result files are in a comma-separated-value (CSV) format, with the following structure:

  Patient,GenoID,0,1,2
  0,17084716,-0.0519483,0.0480828,-0.0553241
  0,17084761,-0.0602524,0.0287036,-0.036127
  0,17085108,0.00387097,-0.0046014,-0.0691283
  0,17086613,0.00317046,-0.00328952,-0.0551729
  

• To reproduce the results obtained for the paper "Secure and Practical Genotype Imputation Methods via Homomorphic Encryption" and re-run the encrypted predictions for the test dataset, run the following commands

  make build
  

  make run REFDATAPATH=prediction_data/original_data/tag_testing_trimed.txt WINDOW=16 ENCGOROUTINES=4 ENCBATCHSIZE=250 PREDGOROUTINES=4 PREDBATCHSIZE=5000 DECGOROUTINES=4 DECBATCHSIZE=2500 NBRTARGETSNPS=80882 NBRPATIENTS=1004 POPTYPE=ALL
  

  • Note
    • REFDATAPATH is the path of input files (tag SNPs)
    • WINDOW is the plaintext model type (equal to the number of neighboring SNPs used for the prediction + 1)
    • NBRPATIENTS and POPTYPE should be consistent with the used dataset (e.g., 1004 for ALL; 272 for AFR; 135 for AMR; 210 for EUR)

  python3 results/transform.py -i results/ypred.binary -t prediction_data/original_data/answer_targets_80882.txt -o results/curve -c window16 -p 1004
  

  • The arguments of the transform.py script are the following:
    • -i : input predictions
    • -t : answer files
    • -o : curve figure
    • -c : csv file name
    • -p : number of patients