Performance of different polygenic risk scores for breast cancer risk prediction: in-depth evaluations across large UK and Australian cohorts

Polygenic risk scores (PGS) have the potential to support enhanced, risk-based screening for breast cancer. Previous studies for many diseases found that genome-wide PGS (GW-PGS) outperform PGS derived by applying hard GWAS significance thresholds. To support future breast cancer risk predictions, we compared the predictive performance of two existing PGS (including PGS313, a leading hard-thresholding PGS) and five newly developed GW-PGS (applying different methods to recent GWAS). We evaluated the performance of PGS Z-scores and of predicted 5-year absolute breast cancer risks based on age alone or age and PGS, across three large cohorts from the UK (UK Biobank) and Australia (QSkin, Melbourne Collaborative Cohort Study). Performance was assessed using discrimination (AUC) and calibration metrics, with dedicated evaluations for European, South Asian and African genetic ancestry groups, different age groups and for UKB, by pre-baseline mammogram screening history. Z-scores from three GW-PGS (LDpred2, PRS-CS, PRS-CS₂₀₁₇) yielded improved discrimination over PGS313, especially in European and South Asian ancestry groups (AUC improvements 2-18%, p < 0.029). Incorporating PGS substantially improved absolute risk predictions compared to age-only models, with the strongest evidence in European-ancestry groups (AUC improvements 15-39%, p < 10⁻⁴) and similar trends in non-European groups. No PGS outperformed all others across all ancestry groups. Estimated relative risk for highest GW-PGS risk groups (e.g. top 5% LDpred2) was ~2.5-fold population-average risk, similar to previous estimates for individuals with pathogenic variants in ATM and CHEK2 genes. These findings support the potential of PGS for risk-based breast cancer screening, noting that current GW-PGS may not substantially improve breast cancer risk predictions compared to PGS313.