Integration of a Cross-Ancestry Polygenic Model With Clinical Risk Factors Improves Breast Cancer Risk Stratification

Abstract

PURPOSE To develop and validate a cross-ancestry integrated risk score (caIRS) that combines a cross-ancestry polygenic risk score (caPRS) with a clinical estimator for breast cancer (BC) risk. We hypothesized that the caIRS is a better predictor of BC risk than clinical risk factors across diverse ancestry groups. METHODS We used diverse retrospective cohort data with longitudinal follow-up to develop a caPRS and integrate it with the Tyrer-Cuzick (T-C) clinical model. We tested the association between the caIRS and BC risk in two validation cohorts including > 130,000 women. We compared model discrimination for 5-year and remaining lifetime BC risk between the caIRS and T-C and assessed how the caIRS would affect screening in the clinic. RESULTS The caIRS outperformed T-C alone for all populations tested in both validation cohorts and contributed significantly to risk prediction beyond T-C. The area under the receiver operating characteristic curve improved from 0.57 to 0.65, and the odds ratio per standard deviation increased from 1.35 (95% CI, 1.27 to 1.43) to 1.79 (95% CI, 1.70 to 1.88) in validation cohort 1 with similar improvements observed in validation cohort 2. We observed the largest gain in positive predictive value using the caIRS in Black/African American women across both validation cohorts, with an approximately two-fold increase and an equivalent negative predictive value as the T-C. In a multivariate, age-adjusted logistic regression model including both caIRS and T-C, caIRS remained significant, indicating that caIRS provides information over T-C alone. CONCLUSION Adding a caPRS to the T-C model improves BC risk stratification for women of multiple ancestries, which could have implications for screening recommendations and prevention.

Addition of polygenic risk score to a risk calculator for prediction of breast cancer in US Black women

“…The addition of the PRS increased the AUC in both models, from 0.56 to 0.59 in the BCRAT and from 0.51 to 0.55 in the IBIS model. Most recently, Tshiaba et al have evaluated the addition of a cross-ancestry PRS [ 40 ] to the IBIS model in data from the Women’s Health Initiative and the UK Biobank [ 41 ]. Across all ancestry groups, the addition of the PRS to the IBIS model increased the AUC for prediction of risk in the next five years from 0.56 to 0.65 in the WHI and from 0.57 to 0.63 in the UK Biobank.…”

“…In summary, by combining estimates from the previously validated BWHS breast cancer risk prediction model with the newly validated AA-PRS, we now have a combined model that provides discriminatory accuracy higher than the BWHS model alone and similar in magnitude to combined models in women of European ancestry. Cross-ancestry models are being put forth as valuable for multiple ancestral populations, but, to date, show relatively poor performance in African ancestry populations [ 41 ]. To develop a cross-ancestry PRS that works well for all major population groups, it will be necessary to have larger numbers of cases and controls from African ancestry populations and from other populations currently underrepresented in genetics research.…”

Recent advances in polygenic scores: translation, equitability, methods and FAIR tools

“…For example, a recent study of blood lipid levels showed that PGS constructed using multi-ancestry GWAS outperforms those constructed using single-ancestry matched data [ 118 ]. A larger analysis of 14 disease endpoints results from the Global Biobank Meta-analysis Initiative (GBMI) also concluded that using multi-ancestry GWAS improved the accuracy of PGS for all ancestries, although a significant amount of heterogeneity in accuracy exists across ancestries [ 119 ], and many other PGS based on multi-ancestry GWAS can be validated in diverse populations [ 46 , 120 , 121 ]. However, multiple studies constructing and evaluating PGS in African populations have come to the opposite conclusion that ancestry-matched PGS is most optimal for the prediction [ 25 , 115 , 122 , 123 ].…”