Modeling Genetic Risk of Breast Cancer

Starting even before the identification of the BRCA1 and BRCA2 cancer susceptibility genes, several models were well established as predictors of a woman's risk of inherited breast cancer.¹ During the past decade, these models have been refined not only to predict breast cancer risk, but also to predict the likelihood that an individual carries a deleterious mutation in one of these genes. Such models have been used as criteria for offering genetic testing,² for determining eligibility for screening and prevention trials,³ and, increasingly, for assessing appropriateness of participation in incremental risk-reduction strategies.^{4 - 5} Illustrating the influence modeling is having on clinical practice, the American Cancer Society recently issued guidelines for offering screening breast magnetic resonance imaging, which, among other criteria, stated that women with a lifetime risk of breast cancer of “20-25% or greater, as determined by BRCAPRO or other models that are largely dependent on family history,” should be offered annual breast magnetic resonance imaging beginning at age 30 years.⁴

In this issue of JAMA, Weitzel and colleagues⁶ present results of a study that details an important limitation of genetic risk assessment modeling. In this study, the authors evaluated the performance of 3 commonly used risk assessment models in 306 women who were diagnosed as having breast cancer prior to age 50 years and had no first- or second-degree relatives with breast or ovarian cancer. This study cohort represents an important fraction of individuals presenting for genetic counseling and testing and accounted for 19.8% of the 1543 individuals presenting to the study center in a 10-year period. In this cohort of women with apparently sporadic, early onset breast cancer, 9.5% of study participants had a deleterious BRCA1 or BRCA2 mutation detected on germline screening.

The authors hypothesized that given the strongly sex-limited penetrance of mutations in BRCA1 and BRCA2, current risk assessment models might not adequately predict mutation probability in these nonhereditary-appearing cases. Such a hypothesis recognizes that a dominant susceptibility trait, such as observed with BRCA1- and BRCA2-linked breast cancer, may be obscured because of a paucity of female relatives in either the maternal or paternal lineage. The authors found that limited family structure (≥1 lineage with <2 first- or second-degree female relatives older than 45 years), compared with adequate family structure (≥2 first- or second-degree female relatives older than 45 years in both the maternal and paternal lineage), was strongly associated with the presence of a deleterious mutation (odds ratio, 2.8; 95% confidence interval, 1.19-6.73; P = .02). Despite this, none of the models evaluated (Myriad,⁷ Couch,⁸ or BRCAPRO⁹ ) demonstrated significantly different mean pretest mutation probabilities when the families were stratified as having limited or adequate family structure.

The authors then evaluated the performance of these 3 risk assessment models as well as a model that tested all individuals with limited family structure by examining the area under the receiver operating characteristic curves. This assessment demonstrated that even though both BRCAPRO and the limited family structure model performed better than chance, neither model performed particularly well. Limiting the analysis to the 245 non–Ashkenazi Jewish participants, the group for which modeling is most needed, BRCAPRO identified only 5 (23.8%) of the 21 mutation carriers in the cohort as having a greater than a 10% probability of a BRCA1 or BRCA2 mutation. The limited family structure model had better sensitivity, identifying 14 (66.7%) of the 21 mutation carriers, yet it achieved this sensitivity at the cost of specificity by recommending testing for 48% of participants in the cohort. Although a combined model created by the authors, incorporating both BRCAPRO probabilities and information regarding adequacy of family structure, appeared to perform better than any of the other models, review of the relevant receiver operating characteristic curve reveals that 90% sensitivity is still achieved with less than 50% specificity.

Given these limitations, what is the role of risk assessment models in individuals with isolated early onset breast cancer? Weitzel and colleagues make a compelling argument that risk assessment models are likely not appropriate if only a limited number of informative relatives are available in either the maternal or paternal lineage. Perhaps more important than this specific conclusion are the implications this study has for use and interpretation of risk assessment models in general. The authors clearly demonstrate that in this cohort, several of the most commonly used risk assessment models overestimate mutation probability, and the resultant lifetime cancer risk, in the setting of adequate family structure. These same models underestimated the risk in the setting of less informative family structure. Given these limitations, important questions are whether currently available models are appropriate to triage individuals for genetic testing or are adequate to provide cancer risk prediction and guidance of care in the absence of genetic testing.

Model-driven estimates clearly have a role as a component of the process of determining the appropriateness of genetic testing for an individual patient. However, this decision process needs to reflect variables that current risk assessment models do not fully assess, including the “informativeness” of the family history (with regard to Mendelian transmission), pathologic and immunohistochemical features of tumors that may be suggestive of inherited predisposition and predict outcome (eg, “triple-negative” breast cancer phenotype),¹⁰ and presence in the kindred of other, less common cancers that may be part of the BRCA mutation spectrum. For these reasons, strict reliance on a single point estimate from a risk assessment algorithm as a threshold to inform decisions regarding appropriateness of genetic testing should be discouraged. Instead, such thresholds probably should be avoided, and, where possible, clinical factors should be taken into account, as proposed in updated guidelines from the American Society of Clinical Oncology.¹¹ These guidelines recommend that genetic testing should be offered if (1) the individual has personal or family history features suggestive of a genetic cancer susceptibility condition; (2) the test can be adequately interpreted; and (3) the test result will influence medical management.¹¹

When applying quantitative risk assessment models to determine appropriateness of intensive cancer screening or risk-reducing surgical approaches, even more caution is indicated. Most current quantitative risk assessment models determine breast cancer risk by creating a weighted estimate that incorporates the probability that a woman will develop breast cancer if she has a mutation in BRCA1 or BRCA2 and the probability that a woman will develop breast cancer if she does not have a mutation in one of these genes. None of the models evaluated in the report by Weitzel et al take into account the possibility that breast cancer risk may be conferred by an as yet unidentified cancer susceptibility gene, despite strong evidence that almost half of hereditary breast cancer is caused by genes not linked to BRCA1 or BRCA2.¹² A more recent model developed by Tyrer et al¹³ does take into account the possibility of a single putative low-penetrance cancer susceptibility gene. However, there is some concern that this model may overestimate cancer risks by incorporating epidemiologic and hormonal factors that may influence the risk of sporadic breast cancer but have not been validated as risk factors for inherited breast cancer.

Of additional concern, no current breast cancer risk prediction model takes into account the genetic and allelic heterogeneity and phenotypic diversity of inherited breast cancer. For example, BRCA1 mutation carriers are at greatly increased risk of both estrogen receptor–negative breast cancer^{14 - 15} and premenopausal ovarian cancer,^{16 - 17} whereas BRCA2 mutation carriers are predominately at risk of estrogen receptor–positive breast cancer and postmenopausal ovarian cancer.^{14 - 18} Preliminary data have also suggested that women from BRCA-negative hereditary breast cancer families may not be at increased risk of ovarian cancer.¹⁹ Given that appropriate screening and primary prevention strategies will likely differ between these distinct syndromes and that no currently available risk assessment model has adequate discriminatory ability to differentiate between these syndromes, it is not clear that present risk assessment models are ready to be used as a surrogate for genetic testing.

Predictive models derived from empirical clinical data have been used in almost every field of medicine and clearly have a role in the risk assessment of breast cancer, the most common malignancy affecting women.²⁰ However, these models are tools that are not intended to be used in isolation, and they have limitations inherent in their specific designs. Studies such as that by Weitzel and colleagues that highlight the limitations of these predictive tools are critical to understand how best to incorporate quantitative models more effectively into the practice of genetic risk assessment and preventive medicine.