1. Introduction
Breast cancer (BC) is one of the most common cancers in the world, with 2.26 million new cases, and killing over 684,000 women in 2020 [
1]. The disease has seen great reductions in recent decades [
2] due to the increasing number of medical therapies, use of screening methods, and emphasis on preventative efforts [
3,
4]. Already, in both the United States and the Netherlands, BC screening begins at age 50 for average risk women, which has led to important decreases in BC-related mortality in both countries [
3,
5]. However, reductions in cancer incidence have slowed in multiple domains, especially those which have improved outcomes from early detection, such as BC [
6]. Increasing preventative efforts has effectively decreased both the morbidity and mortality of cancers; however, addressing the underlying risk factors which lead to BC is still needed [
7].
Enhancing the methods of BC prevention would play a significant role in reducing the number of women who will develop cancer in their lifetimes. One of the major risk factors in the development of BC is an individual’s genetic code. Many advances have been made to characterize the genetic underpinnings of BC [
8], most notably the discovery of BRCA1 [
9], CHEK2 [
10], and ATM mutations [
11]. While the identification of these genes has played a major role in understanding those at very high risk of familial BC, most women who develop BC do not possess these mutations [
8]. In fact, approximately 0.12% of Caucasian women actually carry susceptible variants in BRCA1, accounting for only 1.7% of BC cases diagnosed before age 70 [
9]. Heritability estimates of BC are between 15% and 31% [
12,
13,
14], leaving considerable potential for other genetic variants to contribute to overall genetic risk for BC.
Polygenic risk scores (PRSs) have become an increasingly popular modality for the use of genetics in medicine, particularly for the early detection of individuals with high genetic risk for a given disease [
15]. With the steady expansion of genome-wide association studies (GWASs), genetic data on multiple diseases are now readily available for PRS production. PRSs are developed through the use of GWAS data to determine the single nucleotide polymorphisms (SNPs) associated with the development of a given disease or the phenotype of interest. Specifically, the number of risk alleles at each locus is determined (either 0, 1, or 2) and then weighted by the associated β-value (a measure of relative risk of the effect allele vs. the reference allele) from the discovery GWAS. While the effect of one SNP is generally quite small, by using an additive model and summing the effects of all contributing SNPs in the DNA of an individual, a composite score can be produced and then compared to a reference population for relative genetic risk determination. This method has the potential to scale far more broadly across a population because it does not rely on single-gene testing and can be applied, at a relatively low cost, as a screening and disease prevention tool. Others have demonstrated a significant improvement to the development of PRS algorithms and the potential benefits of their use for categorizing individuals into specific risk categories [
16]. By doing so, patients would be given more personalized advice on lifestyle choices, screening methods, and treatment options to prevent or manage a disease [
16].
In a previously published study by Mavaddat et al. in 2019, the authors developed BC PRSs optimized for the prediction of estrogen receptor-specific BC, utilizing a large collaborative effort involving 94,075 BC positive cases and 75,017 BC negative controls of European ancestry genotyped on the iCOGS and OncoArray following 1000 genomes imputation [
17]. The PRSs were independently and prospectively validated in a test dataset comprising 11,428 breast cancer-affected cases and 18,323 controls. The two best PRSs, 313-SNP and 3820-SNP PRSs, were developed using hard-thresholding, least absolute shrinkage, and selection operator (LASSO) methods, respectively. Validation testing revealed area under the receiver operating characteristic curve (AUROC) values of 0.639 and 0.646 for overall BC for the 313-SNP and 3820-SNP PRSs, respectively. When tested prospectively, the AUROC values decreased slightly to 0.630 and 0.636 for overall BC, respectively, for the 313-SNP and 3820-SNP PRSs. Though the AUROC values were modest, the differences in the tails of the distribution of predicted risk were large, as demonstrated by women in the top 1% of the distribution having a predicted risk approximately four-fold larger than the risk of those in the middle quintile. Furthermore, the association between PRS and disease risk was observed for women with and without a family history of disease.
The clinical potential of PRSs faces a significant hurdle related to the diverse genotyping methods used to procure the necessary genetic data. When essential variants crucial for PRS calculations are not directly identified by genotyping platforms, genotype imputation becomes essential. This process leverages high linkage disequilibrium (LD) genetic variants to infer untagged SNPs, reducing costs compared to whole-genome sequencing. While imputation has been successful, it becomes less precise with rare alleles, potentially yielding slightly different genotypes. These variations, especially at critical loci, may alter an individual’s overall genetic risk, impacting the resulting PRS and overall risk categorization [
18]. Such variability raises concerns regarding proper risk stratification and subsequent clinical care. Here, we assess the variability in BC PRS due to the imputation process.
This study aims to replicate and assess the efficacy of two previously established PRSs within a cohort of BC cases and cancer-negative controls. Additionally, we aimed to assess the impact of genotype imputation on BC PRSs. We anticipate that these PRSs will effectively gauge BC risk among our study subjects, showing comparative performance to prior findings. Additionally, we expect slight variations in BC PRSs due to imputation disparities and the inherent probabilistic nature of genotype inference from reference populations. Through validation across two independent European ancestry populations, this research aims to bolster the evidence supporting PRSs’ utility in differentiating high- and low-risk individuals, facilitating improved stratification for screening and preventative therapies while optimizing risk–benefit ratios. Furthermore, our findings offer insights into the influence of genotype imputation on PRSs, presenting potential strategies for addressing associated concerns.
4. Discussion
Significant progress in genotyping technology has allowed for the accrual of large amounts of genetic data around the world. Over the last decade, researchers have developed methods for making PRSs for various diseases which have the potential to accurately determine the genetic risk of an individual for a given disease and to help improve disease prevention efforts. In this study, our goal was to validate the use of two PRSs for BC in representative case and control populations and to examine the effect of genotype imputation on PRS. Overall, our investigation into PRS models for BC prediction revealed intriguing insights, shedding light on the nuanced interplay between genetic predisposition, population characteristics, and predictive accuracy across different cohorts of similar genetic ancestry.
Based on our analyses, we can confirm that both the 313-SNP and 3820-SNP PRSs perform moderately well in distinguishing individuals with and without BC. We used PCA to generate standardized residual scores for reliable comparisons and used age as a covariate to ensure that the limited allele frequency differences in the case and control cohorts were corrected for and to account for potential confounding. Our data showed a significant difference between the case and control groups for both the 313-SNP and 3820-SNP PRS models (p < 0.01), with the case mean PRSs being significantly greater than the mean PRSs of the respective control groups, confirming the capacity for the 313-SNP and 3820-SNP PRS models to distinguish BC cases from controls.
Our analysis of the 313-SNP and 3820-SNP PRS models further delineated their performance across populations. Notably, both PRS models exhibited significant predictive power for BC, with the 3820-SNP PRS demonstrating superior performance in both study cohorts. This observation was consistent across the overall study population and within individual countries.
The logistic regression models emphasized the substantial enhancement in Nagelkerke R2 values upon inclusion of both PRS models, underlining their utility in BC prediction. Additionally, the area under the curve (AUC) values derived from receiver operator characteristic (ROC) analysis showcased moderate but consistent improvements in predictive performance for the 3820-SNP PRS compared to the 313-SNP PRS, across both American and Dutch populations.
A noteworthy finding was the disparity in prediction accuracy between the American and Dutch cohorts. Our study indicated a higher predictive value of PRS for BC in the American samples compared to those obtained from the Netherlands. We highlight potential factors that may explain this discrepancy. Firstly, the genetic proximity of the American samples to the original GWAS population and study by Mavaddat et al. [
17] potentially conferred a higher predictive accuracy. Secondly, we identified seven variants exhibiting significant allele frequency differences between the Dutch and American samples, which unveiled a plausible reason for the divergent predictive abilities in the two populations. These variants possessed the largest allele frequency differences between cases from the two populations and potentially underpin the enhanced predictive value observed in the American population. The identified variants warrant further investigation to elucidate their functional relevance in breast cancer susceptibility, potentially enriching the predictive power of future PRS models for specific populations. Finally, the distinct nature of case identification—diagnosed breast cancer cases in the USA based on medical records versus population-based surveys in the Netherlands—might contribute to the varying severity levels captured within the cohorts, potentially impacting the overall predictive accuracy of the PRS models.
Moreover, our investigation into the impact of genotype imputation on BC PRSs revealed noteworthy findings. Through 20 imputation replications, we observed a 2–3% variation in resultant BC PRS, aligning with previous findings [
18]. The study demonstrated that utilizing different pre-phasing and imputation tools resulted in minimal percentile changes (<5%) across 14 PRSs, encompassing different disease architectures and PRS calculation approaches. However, in our study, there were also people with substantially larger differences, even reaching up to 9%. This highlights the challenges in individual-level genetic analysis where rare variability events can be obscured by high overall score reproducibility at the population level. To address potential fluctuations in the PRS results of some individuals, our suggestion echoes that of Chen et al. [
18], in that our recommendation would be to employ the average of multiple imputation iterations when calculating PRSs for clinical use and or personal predictions, when using imputed genotype data. This will mitigate the stochastic nature of inferring alleles, ensuring more robust and consistent PRS outcomes.
Understanding how PRS assesses risk and how this may impact clinical decision making, it is important to differentiate between relative and absolute risk. To do so, we utilized the GenoPred web tool [
29] to convert PRS to relative risk and absolute risk for each of our case and control populations. The relative risk differences were approximately 10–15% between cases and controls for the 313-SNP PRS in both the US and NL cohorts. This relative risk difference expanded in the US cohort to nearly 20% and shrank to about 8% in the NL cohort for the 3820-SNP PRS. The absolute risk for the US cohort was 0.40% greater for the average US case (2.60%) vs. the average US control (2.20%) based on the 313-SNP PRS. Compared to the absolute risk reduction (ARR) of 20 years of screening mammograms (0.49%) [
30], the ARR for the 313-SNP PRS of 0.40% is smaller, but not without impact, especially given that the test would only need to be performed once and at an earlier age than mammography. PRS testing may, in fact, be able to improve patient selection for increased mammography frequency or use at an earlier age, as some studies suggest [
31].
This ARR is increased as the PRS increases, as exemplified by a US BC case in our study with a Z-score of 1, the absolute risk is 3.7% (based on 313-SNP PRS AUC), leading to an absolute risk reduction of 1.5% compared to the average US control. As the PRS further increases to a Z-score of 2, the absolute risk becomes 5.50% with a resulting absolute risk reduction of 3.30% compared to the average US control. These calculations suggest that PRS functions to distinguish individuals at the higher end of genetic risk better than individuals closer to the mean. However, these comparisons are difficult to truly assess because PRS functions as a spectrum, rather than a yes–no screening test, leading to lost context.
The results of the percentile bin RR and AR calculations show much smaller differences when comparing similar quartiles between cases and controls, especially for the NL population, where the ARR is often 0.1%. In addition, the combination of our smaller sample size, relatively low disease prevalence, and low AUC likely contributes to these results. Larger prospective studies will likely improve the field’s ability to address these questions.
While our study yields valuable insights into BC PRS modeling, acknowledging limitations is crucial. It is imperative to note that our PRS analyses were conducted after excluding certain SNPs from the 313-SNP and 3820-SNP PRSs. Whenever feasible, direct bi-allelic tagging SNPs were used as replacements, but nevertheless resulted in PRS calculations based on 307 out of the 313 (98.1%) and 3712 out of the 3820 (97.2%) SNPs initially considered. Additionally, the use of population-based surveys for cases in the Netherlands and the potential influences of unaccounted genetic or environmental factors may have contributed to the diminished predictive performance compared to the clinically diagnosed BC cases from the US. The data collected in this study did not contain sufficient information on all cases regarding cancer stage or histochemical subtype, limiting our ability to investigate these characteristics in our study population. This discrepancy highlights the need for ongoing research to refine and enhance the accuracy of PRS models tailored to specific subtypes of disease and diverse populations more generally. Additionally, it is vital to recognize that PRSs are not entirely specific to predicting only the intended trait. They may also forecast other pertinent disease comorbidities or genetically correlated traits. For instance, a documented positive genetic correlation exists between BC and schizophrenia (rg = 0.14, se = 0.03) [
32]. Therefore, variations in PRS predictive performance could stem from prevalence differences of genetically correlated traits among populations.