Secure Inference on Homomorphically Encrypted Genotype Data with Encrypted Linear Models.

Area of Science:

• Genomics
• Computational Biology
• Cryptography

Background:

• Accurate phenotype prediction from high-throughput genomic data is essential.
• Protecting sensitive genetic and phenotypic information requires secure modeling approaches.
• Developing accurate and secure models is critical for privacy-preserving phenotype inference.

Approach:

• A secure inference protocol was developed using homomorphically encrypted genotype data and encrypted linear models.
• Genotype data was scaled by feature importance (XGBoost/Adaboost) and linear models were trained for phenotype prediction.
• The Cryptographic Computation Kit (CKKS) scheme was employed for encrypting model parameters and test data, enabling secure, homomorphic computation and subsequent decryption for evaluation.

Key Points:

• The protocol demonstrated high accuracy, achieving 1-NRMSE values up to 0.9673 for continuous phenotypes and AUC values up to 0.9943 for categorical phenotypes.
• Validation on 5 phenotypes across 3000 samples showed robust performance under random sampling.
• The secure inference achieved an average accuracy of 0.9725 on an encrypted test set, with an overall inference time of 4.32 seconds.

Conclusions:

• An accurate and secure protocol for phenotype prediction from genotype data has been successfully developed.
• The protocol efficiently provides hundreds of phenotype predictions within seconds, ensuring data privacy.
• This method ranked first in the iDASH-2022 track2 challenge, highlighting its effectiveness and speed.