Genome identity matching falters when genuine matches are rare
A peer-reviewed computational study tested whether predicted traits could identify the owner of an anonymized genome. Strong matching performance in a balanced benchmark fell sharply when genuine matches became rare. The UK Biobank analysis examines a particular privacy threat: success depended on the population searched as well as the accuracy of genetic predictions.
Science··Evening
Rare matches change the identification result
Identifying a person from an anonymized genome became much less reliable when genuine matches were rare in a peer-reviewed computational study published on 25 October. Researchers compared genetically predicted traits with observed characteristics, testing a possible route from an unnamed genome back to its owner. The same central finding is covered by Bioengineer.org: a technique that performs strongly under favourable assumptions loses precision under realistic search conditions. The experiments used known genotype–trait pairings from UK Biobank, a British health research resource. Participants were white British, restricting how directly these results apply to populations with other ancestry backgrounds.[1], [2]
Trait predictions were tested in different information settings
The model, called PGMatch, assessed how closely a genome’s predicted characteristics resembled those of a candidate person. Researchers started with 76 traits and retained 59 after removing highly correlated measurements. Their main matching analysis used 40 traits with stronger predictive performance. Polygenic predictions combine information from many genetic variants associated with a characteristic, rather than relying on one gene. The team compared versions trained on individual-level information with versions using summary statistics. Genetic and environmental correlations between traits were included in the assessment, because several related measurements do not provide the same information as wholly independent observations. The comparison also examined the information available to a potential attacker, distinguishing access to paired individual data from aggregate genetic associations.[1]
False matches dominate a large population search
When matching and nonmatching pairs were equally common, precision exceeded 99 per cent while the model recovered 40 per cent of true matches. With a very low probability that the target was present, precision fell below 0.13 per cent at the same recovery rate. Precision is the share of proposed matches that are correct: rare genuine matches allowed false positives to dominate. Population size and prior information therefore affected the calculation alongside predictive accuracy. The analysis concerns this trait-based matching strategy. Genealogical searches and searches already narrowed by additional information fall outside its overall safety claim; the study does not establish general anonymity for shared genomes.[1]
Related columns
For more information on this topic, you can read the related columns.
