Reconstructing Genotypes in Private Genomic Databases from Genetic Risk Scores.

Area of Science:

• Genomics
• Bioinformatics
• Data Security

Background:

• Large genomic databases are frequently re-used for multiple genome-wide association studies (GWAS) and trait investigations.
• Genetic risk score (GRS) models are commonly reported for each trait within publications.

Purpose of the Study:

• To demonstrate that GRS models can be exploited to reconstruct individual genetic variants from genomic databases.
• To identify conditions under which such reconstruction attacks are feasible.

Main Methods:

• Theoretical analysis of GRS model properties.
• Experimental validation using the Cornell Dog Genome database.
• Assessing the impact of aggregate co-occurrence information on attack accuracy.

Main Results:

• Under specific circumstances, particularly with overlapping participant sets between studies, GRS models can enable the recovery of individual genotypes.
• The accuracy of reconstruction attacks is linked to the precision of estimating single nucleotide polymorphism co-occurrence rates within the database.

Conclusions:

• Re-using genomic databases for multiple analyses, especially with partially excluded or included participant subsets, poses a significant privacy risk.
• Releasing aggregate genetic information, such as SNP co-occurrence rates, can severely compromise the security of private genomic databases.