Evaluating anonymized genome re-identification using polygenic predictions and its implications for data privacy

Re-identification by phenotypic prediction aims to determine whether a genome belongs to a specific individual by comparing the individual’s known traits with those predicted from the genome. Although this tracing attack is widely discussed in the genomic privacy literature, its practical risk remains uncertain. As genome-wide association studies (GWAS) continue to improve phenotypic prediction, concerns about the potential for this attack increase. Here, to quantify its real-world threat, we develop a probabilistic framework that estimates the likelihood of a match between an individual’s observed traits and polygenic scores (PGS) derived from a genome, while accounting for prediction accuracy and genetic and environmental correlations between traits. We benchmark this re-identification method and examine how the prior probability that a random genome and set of traits correspond to the same person affects performance. Finally, we assess whether sensitive genetic information, such as APOE- ε 4 status (linked with Alzheimer’s disease) can be inferred through this attack. Our method outperforms a state-of-the-art tool, reaching a precision above 99% for a recall of 40% when considering a prior of 50%. However, in real-world settings, we estimate that realistic priors do not exceed 4 × 10 −4 %, resulting in a precision lower than 0.13% at the same recall. Inference of sensitive genetic variants also proves ineffective: precision above 50% for identifying APOE- ε 4 carriers is achieved only at a recall below 20%. Thus, although re-identification by phenotypic prediction is technically feasible, its effectiveness in real-world conditions remains limited.