Development and Validation of a Breast Cancer Polygenic Risk Score on the Basis of Genetic Ancestry Composition

Abstract

Purpose: Polygenic risk scores (PRSs) for breast cancer (BC) risk stratification have been developed primarily in women of European ancestry. Their application to women of non-European ancestry has lagged because of the lack of a formal approach to incorporate genetic ancestry and ancestry-dependent variant frequencies and effect sizes. Here, we propose a multiple-ancestry PRS (MA-PRS) that addresses these issues and may be useful in the development of equitable PRSs across other cancers and common diseases.

Materials and methods: Women referred for hereditary cancer testing were divided into consecutive cohorts for development (n = 189,230) and for independent validation (n = 89,126). Individual genetic composition as fractions of three reference ancestries (African, East Asian, and European) was determined from ancestry-informative single-nucleotide polymorphisms. The MA-PRS is a combination of three ancestry-specific PRSs on the basis of genetic ancestral composition. Stratification of risk was evaluated by multivariable logistic regression models controlling for family cancer history. Goodness-of-fit analysis compared expected with observed relative risks by quantiles of the MA-PRS distribution.

Results: In independent validation, the MA-PRS was significantly associated with BC risk in the full cohort (odds ratio, 1.43; 95% CI, 1.40 to 1.46; P = 8.6 × 10^-308) and within each major ancestry. The top decile of the MA-PRS consistently identified patients with two-fold increased risk of developing BC. Goodness-of-fit tests showed that the MA-PRS was well calibrated and predicted BC risk accurately in the tails of the distribution for both European and non-European women.

Conclusion: The MA-PRS uses genetic ancestral composition to expand the utility of polygenic risk prediction to non-European women. Inclusion of genetic ancestry in polygenic risk prediction presents an opportunity for more personalized treatment decisions for women of varying and mixed ancestries.

FIG 1.

Summary of independent study cohorts.

FIG 2.

Distribution of PRS in unaffected women from the validation study. (A) Distribution of an 86-SNP PRS, using European-derived weights and allele frequencies applied to women of different ancestries. (B) Distribution of the MA-PRS, using ancestry-specific weights and allele frequencies, in women of different ancestries. (C) Distribution of the MA-PRS in Hispanic women with and without a protective Amerindian SNP. MA-PRS, multiple-ancestry PRS; PRS, polygenic risk score; SNP, single-nucleotide polymorphism.

FIG 3.

Goodness of fit of MA-PRS relative risk. Relative risk of BC, predicted by the continuous MA-PRS (theoretical) or the average relative risk in categories binned by the indicated percentiles (observed), is graphed against the percentiles of the MA-PRS distribution for (A) all ancestries, (B) White and/or Ashkenazi ancestries, and (C) non-European ancestries. ORs are for different percentile categories of the PRS relative to the median category. ORs and 95% CIs are shown. BC, breast cancer; MA-PRS, multiple-ancestry PRS; OR, odds ratio; PRS, polygenic risk score.