Re-identification by phenotypic prediction is a widely discussed threat in the genomic privacy literature. The idea is simple in principle: if genome-wide association studies keep improving the prediction of observable traits from DNA, then someone holding both an anonymized genome and a dossier of an individual’s measurable characteristics could, in theory, link the two. As prediction accuracy rises with ever-larger GWAS, concerns about this so-called tracing attack have grown, yet its practical risk has remained uncertain.

To measure that risk, the researchers built a framework that estimates the likelihood that an individual’s observed traits match the polygenic scores derived from a genome, while explicitly accounting for prediction accuracy and for the genetic and environmental correlations between traits. They benchmarked this re-identification method against a state-of-the-art tool and examined how the prior probability that a random genome and a given set of traits belong to the same person affects performance.

Under favorable assumptions, the method performs impressively. When the prior probability of a correct match was set at 50 percent, the framework outperformed the existing tool, achieving a precision above 99 percent at a recall of 40 percent. In other words, when the attacker already has strong reason to believe a genome and trait set belong to the same person, the approach can confirm matches with very few false positives.

The picture changes dramatically under realistic conditions. The researchers estimated that real-world priors do not exceed 4 × 10⁻⁴ percent, reflecting the enormous number of candidate individuals in any practical search space. At that prior, precision falls below 0.13 percent at the same 40 percent recall, meaning the overwhelming majority of proposed matches would be wrong. The attack’s apparent power, the study shows, depends almost entirely on how much the attacker already knows before starting.

The team also tested whether the attack could expose sensitive genetic information rather than identities. They assessed whether APOE-ε4 carrier status, a variant linked with Alzheimer’s disease, could be inferred through the same trait-matching approach. That inference also proved ineffective: precision above 50 percent for identifying APOE-ε4 carriers was achieved only at a recall below 20 percent, meaning the method could flag only a small fraction of carriers before its error rate became unacceptable.

The findings suggest that although improving polygenic prediction will continue to sharpen the theoretical edge of re-identification attacks, the practical barrier lies in the sheer size of candidate pools and the uncertainty of trait predictions. For now, the authors’ quantitative assessment indicates that anonymized genomes face a limited real-world threat from this vector, a result that may help calibrate ongoing debates about genomic data sharing and privacy safeguards.

Subject of Research: Quantifying the real-world risk of re-identifying anonymized genomes through polygenic trait prediction.

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

Article References: Cavinato, T., Hofmeister, R. J., & Kutalik, Z. (2026). Evaluating anonymized genome re-identification using polygenic predictions and its implications for data privacy. Nature Communications. https://doi.org/10.1038/s41467-026-78494-w

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78494-w

Keywords: genomic privacy, re-identification, polygenic scores, GWAS, phenotypic prediction, APOE-ε4, Alzheimer's disease, data privacy, Nature Communications, bioinformatics, Evaluating, anonymized