A new statistic and its power to infer membership in a genome-wide association study using genotype frequencies

Abstract

Aggregate results from genome-wide association studies (GWAS), such as genotype frequencies for cases and controls, were until recently often made available on public websites because they were thought to disclose negligible information concerning an individual's participation in a study. Homer et al. recently suggested that a method for forensic detection of an individual's contribution to an admixed DNA sample could be applied to aggregate GWAS data. Using a likelihood-based statistical framework, we developed an improved statistic that uses genotype frequencies and individual genotypes to infer whether a specific individual or any close relatives participated in the GWAS and, if so, what the participant's phenotype status is. Our statistic compares the logarithm of genotype frequencies, in contrast to that of Homer et al., which is based on differences in either SNP probe intensity or allele frequencies. We derive the theoretical power of our test statistics and explore the empirical performance in scenarios with varying numbers of randomly chosen or top-associated SNPs.

Figure 1. Histogram of T_geno for a GWAS with 1,000 cases and controls

The figure presents data using 1,000 cases (group 1 in red), 1,000 controls (group 2 in blue) and 1,000 subjects not in the study based on genotypes from Illumina HumanHap550 assay. The theoretical null density curve is shown in black.

Figure 2. Histograms of calibrated T_geno and Homer’s T_allele with 1,000 cases and controls and varying numbers of SNPs

The figure presents theoretical null density curves (black) for a GWAS with 1,000 cases (group 1 in red), 1,000 controls (group 2 in blue) and 12,000 subjects not in the study (in gray) using genotypes for (a) 10,000, (b) 100,000, and (c) 550,000 top associated SNPs from the Illumina HumanHap550 assay. Statistics were calibrated so that the null distribution was centered at zero with unit variance.

Figure 3. Sensitivity and specificity of T_geno applied to GWAS data

Log-scale Receiver Operating Characteristic (ROC) curves of T_geno with Illumina HumanHap550 data from GWAS scenarios with 1000/1000 and 5000/5000 cases and controls of European descent.