Assessing and mitigating privacy risk of sparse, noisy genotypes by local alignment to haplotype databases

Abstract

Single nucleotide polymorphisms (SNPs) from omics data carry a high risk of reidentification for individuals and their relatives. While the ability of thousands of SNPs (especially rare ones) to identify individuals has been repeatedly demonstrated, the ready availability of small sets of noisy genotypes - such as from environmental DNA samples or functional genomics data - motivated us to quantify their informativeness. Here, we present a computational tool suite, PLIGHT ("Privacy Leakage by Inference across Genotypic HMM Trajectories"), that employs population-genetics-based Hidden Markov Models of recombination and mutation to find piecewise alignment of small, noisy query SNP sets to a reference haplotype database. We explore cases where query individuals are either known to be in a database, or not, and consider a variety of queries, including simulated genotype "mosaics" (composites from 2 source individuals) and genotypes from swabs of coffee cups from a known individual. Using PLIGHT on a database with ~5,000 haplotypes, we find for common, noise-free SNPs that only ten are sufficient to identify individuals, ~20 can identify both components in two-individual simulated mosaics, and 20-30 can identify first-order relatives (parents, children, and siblings). Using noisy coffee-cup-derived SNPs, PLIGHT identifies an individual (within the database) using ~30 SNPs. Moreover, even when the individual is not in the database, local genotype matches allow for some phenotypic information leakage based on coarse-grained GWAS SNP imputation and polygenic risk scores. Overall, PLIGHT maximizes the identifying information content of sparse SNP sets through exact or partial matches to databases. Finally, by quantifying such privacy attacks, PLIGHT helps determine the value of selectively sanitizing released SNPs without explicit assumptions about underlying population membership or allele frequencies. To make this practical, we provide a sanitization tool to remove the most identifying SNPs from a query set.