Fast and Scalable Private Genotype Imputation Using Machine Learning and Partially Homomorphic Encryption.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genome sequencing advances offer insights into genetic variation and disease.
• Genotype imputation is crucial for genome-wide association studies but faces privacy challenges due to outsourcing.
• Current machine learning (ML) privacy-preserving methods are not optimized for the large number of outputs required in genotype imputation.

Purpose of the Study:

• To develop a fast, scalable, and accurate privacy-preserving genotype imputation method.
• To adapt linear models for privacy-preserving genotype imputation using homomorphic encryption.
• To evaluate the performance of the proposed method against state-of-the-art plaintext solutions.

Main Methods:

• Utilized Machine Learning (ML) combined with the Paillier cryptosystem for homomorphic encryption.
• Explored the effectiveness of linear models for genotype imputation within a privacy-preserving framework.
• Applied standardized homomorphic encryption schemes to convert linear models for privacy-preserving inference.

Main Results:

• Achieved performance equivalent to state-of-the-art plaintext genotype imputation methods.
• Reached up to 99% micro area under the curve score.
• Demonstrated scalability on large-scale real-world datasets with up to 80,000 targets.

Conclusions:

• The proposed privacy-preserving genotype imputation method is effective and accurate.
• Linear models combined with homomorphic encryption provide a viable solution for secure large-scale genetic data analysis.
• This approach addresses critical privacy concerns in outsourced genotype imputation.