ABSTRACT
Polygenic risk scores (PRS) for breast cancer (BC) have a clear clinical utility in risk prediction. PRS transferability across populations and ancestry groups is hampered by population-specific factors, ultimately leading to differences in variant effects, such as linkage disequilibrium (LD) and differences in variant frequency (AF-diff). Thus, locally-sourced population-based phenotypic and genomic datasets are essential to assess the validity of PRS derived from signals detected across populations. Here, we assess the transferability of a BC PRS composed of 313 risk variants (313-PRS) in a Brazilian tri-hybrid admixed ancestries (European, African and Native American) whole-genome sequenced cohort, GRAR. We computed 313-PRS in GRAR (n=853) using the UK Biobank (UKBB, n=264,307) as reference. We show that although the Brazilian cohorts have a high European (EA) component, with AF-diff and to a lesser extent LD patterns similar to those found in EA populations, the 313-PRS distribution is inflated when compared to that of the UKBB, leading to potential overestimation of PRS-based risk if EA is taken as a standard. Interestingly, we find that case-controls lead to equivalent predictive power when compared to UKBB-EA samples with AUROC values of 0.66-0.62 compared to 0.63 for UKBB.
