Prospective Evaluation of the Addition of Polygenic Risk Scores to Breast Cancer Risk Models

Abstract

Abstract Background The Breast and Ovarian Analysis of Disease Incidence and Carrier Estimation Algorithm and the International Breast Cancer Intervention Study breast cancer risk models are used to provide advice on screening intervals and chemoprevention. We evaluated the performance of these models, which now incorporate polygenic risk scores (PRSs), using a prospective cohort study. Methods We used a case-cohort design, involving women in the Melbourne Collaborative Cohort Study aged 50-75 years when surveyed in 2003-2007, of whom 408 had a first primary breast cancer diagnosed within 10 years (cases), and 2783 were from the subcohort. Ten-year risks were calculated based on lifestyle factors, family history data, and a 313-variant PRS. Discrimination was assessed using a C-statistic compared with 0.50 and calibration using the ratio of expected to observed number of cases (E/O). Results When the PRS was added to models with lifestyle factors and family history, the C-statistic (95% confidence interval [CI]) increased from 0.57 (0.54 to 0.60) to 0.62 (0.60 to 0.65) using IBIS and from 0.56 (0.53 to 0.59) to 0.62 (0.59 to 0.64) using BOADICEA. IBIS underpredicted risk (E/O = 0.62, 95% CI = 0.48 to 0.80) for women in the lowest risk category (<1.7%) and overpredicted risk (E/O = 1.40, 95% CI = 1.18 to 1.67) in the highest risk category (≥5%), using the Hosmer-Lemeshow test for calibration in quantiles of risk and a 2-sided P value less than  .001. BOADICEA underpredicted risk (E/O = 0.82, 95% CI = 0.67 to 0.99) in the second highest risk category (3.4%-5%); the Hosmer-Lemeshow test and a 2-sided P value was equal to .02. Conclusions Although the inclusion of a 313 genetic variant PRS doubles discriminatory accuracy (relative to reference 0.50), models with and without this PRS have relatively modest discrimination and might require recalibration before their clinical and wider use are promoted.