Genome-Based Prediction of Breast Cancer Risk in the General Population: A Modeling Study Based on Meta-Analyses of Genetic Associations

Recent genome-wide association studies (GWAS) have identified various novel breast cancer susceptibility variants (1–4). In contrast to BRCA1 and BRCA2, these susceptibility variants mainly have weak effects and contribute to small increases in breast cancer risk individually. It is commonly agreed that testing single susceptibility genes is not useful for prediction of breast cancer risk, but the question remains whether combining susceptibility loci in risk models could accurately identify women with markedly higher breast cancer risks. Genetic risk models could be useful in the general population to select at risk women for screening or preventive interventions, such as intensified mammography or MRI screening, the use of chemopreventive agents (tamoxifen or raloxifene), and surgical interventions such as oophorectomy and mastectomy. The predictive performance of genetic risk models has been investigated in simulation studies (5–8) and in many empirical studies, including studies on bladder, prostate, and breast cancers (9–12).

Whether genetic risk models will potentially be used in clinical or public health practice to select women at increased risk of breast cancer, first and foremost depends on the availability of an intervention that needs to be targeted. Pharoah and colleagues (13) proposed that the starting age of mammography screening can be varied according to women's genetic risks, rather than recommending screening to all women aged 50 years and more. Accordingly, women at higher genetic risk will have their first mammogram at a younger age and women at lower risk at a higher age than average. An alternative strategy is to vary the frequency of mammograms: women at higher risk will undergo screening more frequently than women at lower risk.

Such differentiation of screening programs requires a risk model that can identify women at increased risk among all women in the population (14). For evaluations at the population level, this predictive performance of a test is assessed as the accuracy of tests to discriminate between women who will develop breast cancer and those who will not. The discriminative accuracy is generally expressed as the area under the receiver operating characteristics curve (AUC). Compared with the measures of reclassification that have been proposed (15–17), AUC is an overall measure of the predictive ability and reclassification a measure of the clinical relevance. Assessment of reclassification only is meaningful when clinically established risk thresholds are available.

Several simulation studies have demonstrated that to accurately classify individuals at high risk of disease, a risk model should either at least include a few genetic variants with strong effects or a high number of variants with small effects (6, 8, 18). We investigated the extent to which breast cancer risk in the general population can be predicted using genetic risk models. In a simulation study, we constructed risk models on the basis of currently identified polymorphisms and hypothetical variants to investigate the future potential in terms of AUC. For this purpose, we reviewed all meta-analyses of genetic association studies on breast cancer. Additionally, we investigated the magnitude of the odds ratios (OR) of 1 to 100 hypothetical polymorphisms that would be needed to achieve similar discriminative accuracy as available breast cancer risk prediction models (19–24).

We used a modeling procedure that has been developed and published previously (6), and which has also been used by others (8). Briefly, the procedure creates a data set with information on genotypes and disease status for a population of 10,000 women. The data set is constructed in such a way that the ORs and frequencies of the genotypes and the disease risk match the specified values, which are obtained from the literature. Predicted breast cancer risks are calculated using Bayes' theorem, which states that the posterior odds of breast cancer for each woman is obtained by multiplying the prior odds by the likelihood ratio (LR) of their genotype status on all polymorphisms. The prior odds is calculated from the baseline population breast cancer risk (p) using the formula p/(1 − p). Under the assumption of independent genetic effects, that is, no linkage disequilibrium (LD) between the genetic variants, the LR is obtained by multiplying the LRs of all individual genotypes that are included in the risk model (25). The LRs of the genotypes of each single genetic variant are calculated from a genotype by disease status contingency table (6). This table is constructed from the frequencies and ORs of the genotypes and the population breast cancer risk. The table can also be constructed from allele frequencies and per-allele ORs when Hardy–Weinberg equilibrium is assumed for the distribution of the genotypes. The frequencies and ORs all are specified as study parameters and varied between the simulation scenarios (see below). The posterior odds are converted into breast cancer risks using the formula odds/(1 + odds). Our model does not include gene–gene and gene–environment interaction, because so far there is no strong empirical evidence for this.

The discriminative accuracy is the extent to which test results can discriminate between women who will develop breast cancer and those who will not (26). The AUC is the probability that the test correctly identifies the woman who will develop the disease from a pair of whom 1 will be affected and 1 will remain unaffected, and ranges from 0.5 (total lack of discrimination) to 1.0 (perfect discrimination). The AUC was obtained as the c-statistics by the R function somers2, which is available in the Hmisc library of R software. All simulations were repeated 20 times to obtain robust estimates of the AUC. In each repetition, the OR of each published variant will be obtained as a new random value from the 95% confidence interval (CI), assuming a normal distribution around the point estimate of the OR. All results are presented as averages of the repeated simulations with 95% CIs. Despite that we take a random OR for the published variants at each new simulation, the CIs of the ORs are small because it is theoretically possible to derive the ORs by a formula and due to the fact that a sample size of 10,000 women was used for this calculation. We choose a simulation approach, because the formula is extremely complex when different genetic models (per allele, dominant/recessive, and per genotype) are considered at the same time. Analyses were performed using R software (version 2.6.1; ref. 27).

Three different scenarios were considered. In each scenario, genotypes and breast cancer status were simulated for 10,000 women, assuming a breast cancer lifetime risk of 10%. It is noted that this lifetime risk varies between populations (14), but our primary outcome measure, AUC, is independent of the magnitude of disease risk, that is, modeling a lower or higher risk will give similar results. The first scenario assessed the AUC for a risk model based on all polymorphisms that were investigated in published meta-analyses (see below), and for a risk model based on statistically significant polymorphisms only. The second scenario assessed the expected AUC when 2 to 5 times as many statistically significant polymorphisms with the same distribution of ORs and genotype frequencies would be known. The final scenario investigated the magnitude of the per-allele ORs of 1 to 100 polymorphisms that need to be added to the risk model to obtain AUCs similar to those of available breast cancer risk prediction models. Table 1 shows that the AUC of currently available breast cancer risk models ranges from 0.555 to 0.762. Therefore, we investigated AUC thresholds of 0.70, 0.75, or 0.80. The ORs of hypothetical variants needed to reach these thresholds were obtained for different frequencies of the risk alleles.

PubMed and HuGE Navigator were searched for meta-analyses on genetic association studies on breast cancer published before May 2010. The PubMed search strategy was based on the keywords “breast cancer,” “meta-analysis” in combination with “gene,” “polymorphism,” or “allele.” Meta-analyses were selected if they were based on genetic association studies that applied a case–control design, included women only, focused on breast cancer risk and were written in English. Meta-analyses were excluded when the reported data were reused in a larger study on the same polymorphism. Summary ORs and genotype frequencies were extracted for all genetic models reported in this article, which could refer to per-allele analyses, comparisons of dominant or recessive effects, or comparisons of homozygous and heterozygous carriers with noncarriers. Summary ORs of the total population were extracted, because not all meta-analysis stratified their data for factors such as age or ethnicity. Therefore, all findings will predominantly apply to European and European American women.

For the simulation study, we calculated the AUC for a prediction model based on all polymorphisms and for a model based on statistically significant polymorphisms. For the latter, we selected polymorphisms for which at least 1 comparison yielded a statistically significant result. When multiple comparisons were statistically significant, we preferred the OR of the homozygous/heterozygous comparisons as the comparison of our first choice, the dominant/recessive comparisons as our second choice and the per-allele analysis as our last choice. Statistical significance was based on the nominal P value (P < 0.05) of the OR per comparison. For the statistically nonsignificant polymorphisms that were included in the simulation study, we also preferred the OR and genotypes for homozygous/heterozygous comparisons over the other comparisons. Because the modeling approach assumes independent effects of the polymorphisms in the risk model, we examined whether polymorphisms were in LD. When R² > 0.40, only the polymorphism with the lowest P value was included in the simulation analyses.

Of the 217 retrieved articles from the literature, 107 met the inclusion criteria (Fig. 1). These articles described 199 meta-analyses on 103 polymorphisms in 70 genes. Ninety-six meta-analyses were excluded because the data had been reused in a larger meta-analysis on the same polymorphism and 3 because polymorphisms were in LD. Of the 96 polymorphisms, in 70 genes that were eligible for inclusion in the simulation analysis, 41 showed statistically significant results for at least 1 genetic model. Eleven of the 96 variants and 5 of the 41 statistically significant variants were in slight LD (R² < 0.40) with other variants in the model.

The AUC was 0.68 for the risk model based on all 96 published polymorphisms and 0.67 for the risk model based on the 41 polymorphisms that were significantly associated with breast cancer risk (Table 2). When 2, 3, 4, or 5 times as many polymorphisms with the same distribution of ORs and genotype frequencies would be identified and included, AUCs were 0.73, 0.77, 0.80, and 0.82, respectively. As a comparison, the AUCs for the combinations of the 7 polymorphisms investigated by Gail et al. (5) and Pharoah et al. (13) were both 0.55.

Table 3 shows the magnitude of the ORs and genotype frequencies of genetic variants that would be needed in addition to the original 41 genetic risk variants to obtain AUCs of 0.70, 0.75, and 0.80. The table shows that to achieve an AUC of 0.70, the minimal OR of 5 additional genetic variants should be 1.5 (95% CI: 1.4–1.5) when their risk genotype frequencies are 0.30. To achieve an AUC of 0.75 with 20 to 100 additional genetic variants, the minimal ORs ranged from 1.2 to 2.1 depending on the frequencies of the risk genotypes. These values were 1.3 to 2.7 to achieve an AUC of 0.80.

This study investigated to what degree genetic risk models can predict breast cancer in a general population setting. We estimated that the AUC would be 0.68 when all 96 polymorphisms investigated were included and 0.67 when only statistically significant polymorphisms were considered. These AUC values are comparable to current breast cancer risk prediction models.

Before further interpreting the public health relevance of our findings, 2 methodologic issues need to be disclosed. First, we assumed that genetic variants inherited independently and that the combined effect of the genetic variants on disease risk followed a multiplicative risk model of independent effects (i.e., no statistical interaction terms were included in the model). Although so far no studies have demonstrated gene–gene interactions with breast cancer risk in general populations, it is still possible that these will be discovered in future studies in larger populations. However, gene–gene interactions only improve the breast cancer risk predictions if their effect sizes are substantially high (e.g., OR > 5). When interaction effects are smaller, their effects on the predictive accuracy will be comparable with that of single gene effects, because by definition their frequencies are lower.

Second, we attempted to compare the AUC of the genetic risk models with observed values for available breast cancer risk prediction models, but this comparison should be made with caution. It is important that risk prediction models should be validated in populations that are representative for the population in which the risk model ultimately is applied (28), and that the models address the same time horizon for the risk prediction. We considered application of the risk model in a general population, but many of the available breast cancer risk models were not evaluated in the general population constructed in a similar population. Some have validated risk models in women who have at least 1 affected relative (Table 1), but not in women who have no positive family history. Finally, we modeled lifetime breast cancer risk, where most published models predicted 5-year risks. In principle, this is a valid comparison, because AUC is independent of disease risk, and we assumed that the effects of the polymorphisms do not vary over time, which so far seems to be a reasonable assumption. The comparison may be weakened if the existing models would have had higher AUC when longer follow-up time had been investigated. This is well plausible, but the effect on AUC is unclear because intermediate risk factors may better predict short-term risk of disease and other risk factors may require long follow-up time to demonstrate their effects. Yet, the breast cancer risk models have AUC values between 0.55 and 0.76, and this range is comparable with risk prediction models based on nongenetic risk factors for other diseases (29).

Of the 96 polymorphisms that have been reviewed in meta-analyses or investigated in GWAS, 41 showed statistically significant associations with breast cancer risk (Table 2). These include the 18 polymorphisms that have been identified in recent GWAS (1, 3, 4, 30–32). The heterozygous OR of the polymorphisms identified in GWAS ranged from 0.84 to 1.36. When future GWAS will identify polymorphisms with per-allele ORs around 1.1, theoretically the predictive ability of the genetic risk model can be improved beyond that of existing models. Yet, even such small improvements still require the discovery of hundreds of new variants (6), or the discovery of the true causal variants, which may have stronger effects than the current variants included. Another avenue is to improve breast cancer prediction by combining genetic with nongenetic risk factors. Breast cancer prediction may be further improved when nongenetic risk factors are unrelated to the genetic factors, such as age, lifestyle, and dietary factors, and even the presence of affected relatives, which in nonfamilial breast cancer is unlikely explained by the low-risk susceptibility genes (33). It should not be expected that risk prediction is markedly improved when nongenetic risk factors are potential intermediate factors in a biological pathway linking the genetic factors to breast cancer (29), such as benign breast disease, personal history of breast cancer, and hormonal factors. On the basis of current knowledge of breast cancer risk factors, it is likely that risk prediction models solely based on nongenetic factors will perform better than models based on common single-nucleotide polymorphism (SNP) alone, as ORs for SNPs tend to be smaller than ORs based on nongenetic factors. Investigating the combined predictive performance of genetic and nongenetic factors is of interest to investigate whether available prediction models can be further improved. Yet, also existing prediction models will only be markedly improved when a larger number of susceptibility variants can be added (5).

What level of predictive performance is required for practical implementation of the risk model depends on the intended use. When prediction models are used to make decisions at the individual patient level, higher AUCs are required compared to when models are used to implement population-based prevention or therapeutic strategies. Whether the risk predictions are used as a strategy to determine the age of starting or interval of mammography screening, they will require that the risk model can accurately identify women at increased risk among all women regardless of family history of breast cancer. Other interventions such as MRI screening, the use of chemopreventive agents, lifestyle behavioral changes, and surgical interventions could also be considered. Yet, the more detail that is desired for risk stratification, the better the prediction model should perform. Specifying age of entry in the mammography screening program by year, as suggested by Pharoah and colleagues, requires better predictive performance than specifying age of entry in broader age categories, for example, enter screening at ages 45, 50, or 55 years. What level of AUC is required for this application is a question to be investigated in future modeling studies.

In conclusion, our analyses show that prediction of breast cancer risk based on low susceptibility variants theoretically can achieve similar predictive performance to existing breast cancer risk models, and can even improve prediction of disease when more variants are being discovered. Whether this predictive performance is sufficient for implementation of the risk models in mass prevention programs is ultimately determined by the intended use of the test and the performance of the interventions, weighing the benefits, harms, and costs.

No potential conflicts of interest were disclosed.

This study was financially supported by grants from the Erasmus University Medical Center Rotterdam and by the Center for Medical Systems Biology in the framework of the Netherlands Genomics Initiative (NGI). A.C.J.W. Janssens was additionally sponsored by the Vidi grant of the Netherlands Organisation for Scientific Research (NWO).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.