Integrating Clinical and Polygenic Factors to Predict Breast Cancer Risk in Women Undergoing Genetic Testing

Abstract

Purpose: Screening and prevention decisions for women at increased risk of developing breast cancer depend on genetic and clinical factors to estimate risk and select appropriate interventions. Integration of polygenic risk into clinical breast cancer risk estimators can improve discrimination. However, correlated genetic effects must be incorporated carefully to avoid overestimation of risk.

Materials and methods: A novel Fixed-Stratified method was developed that accounts for confounding when adding a new factor to an established risk model. A combined risk score (CRS) of an 86-single-nucleotide polymorphism polygenic risk score and the Tyrer-Cuzick v7.02 clinical risk estimator was generated with attenuation for confounding by family history. Calibration and discriminatory accuracy of the CRS were evaluated in two independent validation cohorts of women of European ancestry (N = 1,615 and N = 518). Discrimination for remaining lifetime risk was examined by age-adjusted logistic regression. Risk stratification with a 20% risk threshold was compared between CRS and Tyrer-Cuzick in an independent clinical cohort (N = 32,576).

Results: Simulation studies confirmed that the Fixed-Stratified method produced accurate risk estimation across patients with different family history. In both validation studies, CRS and Tyrer-Cuzick were significantly associated with breast cancer. In an analysis with both CRS and Tyrer-Cuzick as predictors of breast cancer, CRS added significant discrimination independent of that captured by Tyrer-Cuzick (P < 10^-11 in validation 1; P < 10^-7 in validation 2). In an independent cohort, 18% of women shifted breast cancer risk categories from their Tyrer-Cuzick-based risk compared with risk estimates by CRS.

Conclusion: Integrating clinical and polygenic factors into a risk model offers more effective risk stratification and supports a personalized genomic approach to breast cancer screening and prevention.

FIG 1.

Summary of independent study cohorts. *Development 2 was previously published by Hughes et al. CRS, combined risk score; DCIS, ductal carcinoma in situ; HBOC, hereditary breast and ovarian cancer; HNPCC, hereditary nonpolyposis colon cancer; LCIS, lobular carcinoma in situ; PRS, polygenic risk score; SNP, single-nucleotide polymorphism.

FIG 2.

Discriminatory accuracy of CRS over Tyrer-Cuzick or PRS alone in validation 1 (A) and validation 2 (B). CRS, Tyrer-Cuzick, and PRS were evaluated separately in terms of likelihood ratio chi-squared test statistics from age-adjusted logistic regression models. In both validation studies, the CRS performed significantly better than either Tyrer-Cuzick or PRS at discriminating between women with and without invasive breast cancer. CRS, combined risk score; PRS, polygenic risk score.

FIG 3.

Evaluation of concordance between CRS (blue) and Tyrer-Cuzick (red). Estimates for average remaining lifetime (A and C) and 5-year risk (B and D) for unaffected controls in validation 1 (A-B) and validation 2 (C-D). Patients were grouped into 5-year age bins, and average risks were evaluated according to CRS and Tyrer-Cuzick. 95% CIs are also reported. CRS, combined risk score.

FIG 4.

Distribution of CRS risk estimates in unaffected women. (A) Remaining lifetime risk (RLR) in a population of unaffected women (clinical performance population, N = 32,576; excluded women with ductal carcinoma in situ, lobular carcinoma in situ, hyperplasia, or unspecified breast disease) according to CRS with thresholds at 20% (increased) and 50% (high) RLR. (B) Scatterplot of RLR based on the Tyrer-Cuzick and CRS risk models for patients within the clinical performance population. (C) Distribution of patients above and below the 20% RLR threshold in the clinical performance population according to both the Tyrer-Cuzick and CRS models. Blue squares indicate patients with discordance between the scores (eg, the Tyrer-Cuzick model produced a score that indicated a patient had low RLR, but the same patient was determined to have increased RLR by the CRS model). CRS, combined risk score.