Ancestry bias in variant effect predictors
Authors/Creators
Description
Variant effect predictors (VEPs) have become increasingly important for clinical variant interpretation. However, many widely used methods are trained using clinically curated or population-derived variant datasets that disproportionately represent individuals of European descent. To test whether this imbalance leads to ancestry-associated differences in predictor output, we evaluated the scoring patterns of 55 missense VEPs across variants from 14 ancestry groups. We observed striking disparities, with several VEPs predicting a markedly higher proportion of damaging variants in underrepresented populations compared to individuals of European ancestry. In contrast, VEPs developed without direct exposure to clinical or human population variants predicted more consistent pathogenicity burdens across populations. These discrepancies closely tracked differences in ancestry representation in variant resources, implicating biased training data as a major contributor. Our findings underscore the urgent need to adopt tools that minimise ancestry bias, ensuring more accurate variant effect prediction and genetic diagnosis for all populations.