Evaluating mammographic density polygenic risk score for contralateral breast cancer risk prediction.

Introduction
Contralateral breast cancer (CBC) is a significant concern for breast cancer survivors, who face a substantially increased risk of developing a second primary malignancy in the opposite breast. Studies indicate that this risk may be two to six times higher than that of the general population developing a first breast cancer1,2. This underscores the urgent need for better predictive models to identify individuals at higher risk of CBC, enabling personalized care and support.
Mammographic density (MD) is a well-established risk factor for first primary breast cancer, with studies demonstrating that women with higher MD have a significantly increased risk of developing the disease. For instance, a meta-analysis by McCormack and dos Santos Silva reported that women with ≥ 75% breast density had a 4.64-fold increased risk compared to those with < 5% density3. Similarly, Boyd et al. found that extensive MD is strongly associated with higher breast cancer risk, with an odds ratio of 4.7 for women with ≥ 75% density compared to those with < 10%4. These findings underscore the importance of MD as a significant factor in breast cancer risk assessment.
While the association between MD and first primary breast cancer is well-documented, its role in CBC risk has been less well-studied. Research from the WECARE Study has provided evidence that higher MD is also associated with an increased risk of CBC5, aligning with the results reported by Sandberg et al.6. Specifically, prior studies have shown that percent MD measured after the initial breast cancer diagnosis may serve as a useful predictor in CBC risk models. These associations appear to be stronger among women diagnosed with their first breast cancer at a younger age, suggesting that MD may have greater predictive value in this subgroup. This highlights the potential of incorporating MD into personalized screening and prevention strategies for breast cancer survivors.
The genetic basis for MD has been studied7,8. MD is influenced by a polygenic architecture, where multiple genetic variants each exert a small effect on the overall phenotype9–11. The heritability of MD-related traits, including dense area (DA), non-dense area, and percentage of dense area (PDA), is substantial, with estimates exceeding 60%12–14.
Genome-wide association studies (GWAS) have identified numerous genetic variants associated with MD15–19. However, the high heritability and the limited variance explained by these GWAS findings suggest that additional genetic variants remain to be discovered. This includes potential gene-gene and gene-environment interactions.
Polygenic risk scores (PRSs) have emerged as a powerful tool to capture the cumulative effects of multiple genetic variants associated with disease risk20. In the context of MD, a PRS may provide a stable, lifelong indicator of a woman’s inherited predisposition to higher MD, complementing measured MD, which can vary with age, treatment, and hormonal exposures. Importantly, a PRS for MD may be particularly valuable when mammograms are unavailable, for example, in younger women, in settings with limited imaging access, or in large-scale genetic studies. Integrating PRS for MD into risk models holds the potential to significantly improve risk stratification, particularly for women who may not fall into conventional high-risk categories based on family history of breast cancer or clinical factors21.
Although feasible, the application of PRS to MD remains understudied, especially in its potential for predicting CBC risk. To address this gap, we leverage a PRS derived from MD-associated genetic variants identified in prior GWAS to evaluate whether incorporating genetic predisposition to MD improves prediction of DA and CBC risk among participants in the WECARE Study.
By advancing our understanding of these relationships, we aimed to contribute valuable insights into the clinical utility of the joint consideration of MD and PRS for MD in personalizing CBC screening and prevention strategies among breast cancer survivors at risk of developing CBC.

Results

Participant characteristics
Table 1 presents the characteristics of the 399 European-ancestry WECARE study participants with mammograms available prior to their first breast cancer diagnosis and who underwent genome-wide genotyping. Most participants were between 45 and 54 years old at the time of their mammogram, with 64% having a BMI of less than 25 kg/m2. Most women were premenopausal (75%) at the time of their mammogram, and 28% reported a first-degree family history of breast cancer.
Among participants, 66% had an estrogen receptor-positive first breast cancer, 65% received chemotherapy, 67% underwent radiation therapy, and 60% received hormone therapy for their first breast cancer. Alcohol consumption and smoking were reported by 70% and 39% of the participants, respectively. Most mammograms were performed between 1995 and 2008, with 94% of the women receiving their mammogram within 12 months prior to their first breast cancer diagnosis.

Mammographic density (MD) measurements
Figure 1 shows the distribution of MD measurements; panel A displays the untransformed DA (in cm²), revealing a right-skewed pattern with most participants having low DAs and a smaller group with higher densities. Panel B presents the measures after log-transformation and standardization, which normalizes the distribution, making it more symmetric and better suited for comparison across MD measurements. This transformation also makes the three MD measures more directly comparable by placing them on the same standardized scale and visually illustrates that Cirrocumulus captures a narrower, higher‑intensity portion of the density distribution compared with Cumulus and Altocumulus. All further analyses used these log-transformed, standardized measures.

Polygenic risk score (PRS) analysis
Of the 64 genome-wide significant variants (P-value <5e-08), 37 were retained for the final PRS_MD calculation. The excluded SNPs were removed due to high linkage disequilibrium (LD) with other variants (r² > 0.01) to minimize redundancy and ensure that the PRS_MD reflected independent genetic contributions. The distribution of the resulting PRS_MD is shown in Fig. 2, where the histogram shows that the PRS_MD is approximately normally distributed in participants with a central peak around zero.

Associations of PRS_MD with MD measurements
The associations between the PRS_MD and MD measures (Cumulus, Altocumulus, and Cirrocumulus) are expressed as log odds per adjusted standard deviation and summarized in Table 2. In multivariable linear regression models, the effect estimate for PRS_MD in relation to Cumulus was 0.62 (95% CI: 0.1,1.1; P = 0.019). The result for Altocumulus was 0.63 (95% CI: 0.1,1.1; P = 0.017). In contrast, the effect estimate for Cirrocumulus was 0.41 (95% CI: -0.1,0.9; P = 0.12), which was not statistically significant.
Table 3 compares the R-squared and adjusted R-squared values for models predicting MD measurements (Cumulus, Altocumulus, and Cirrocumulus) with and without the PRS_MD in participants from the WECARE Study. R-squared values represent the proportion of variance in MD measurements explained by each model, while adjusted R-squared values account for number of predictors, providing a more accurate reflection of model performance.
For the Cumulus model, the model without PRS_MD had an R-squared of 20.6% (adjusted R-squared of 16.9%). When PRS_MD was included, the R-squared increased to 22.8% (17.8%), demonstrating the added predictive value of PRS_MD.
For the Altocumulus model, the model without PRS_MD had an R-squared of 22.7% (19.0%). After including PRS_MD, the R-squared increased to 24.7% (19.9%), also showing a small but notable increase in predictive power with the addition of PRS_MD.
In contrast, for the Cirrocumulus model where the model without PRS_MD had an R-squared of 20.6% (16.9%), including PRS_MD increased the R-squared to 21.8%, but the adjusted R-squared showed a decrease to 16.7%.
These results suggest that PRS_MD contributes to improving the predictive models for Cumulus and Altocumulus mammographic densities but has a more limited impact on Cirrocumulus density.
Full results showing all covariates can be found in Supplementary Tables S1–S3.

Associations of MD measurements and PRS_MD with CBC risk
Table 4 summarizes the associations between MD measurements (Cumulus, Altocumulus, and Cirrocumulus), PRS_MD, and the risk of CBC. Odds ratios (ORs) are from multivariable‑adjusted models and represent the change in risk per adjusted standard deviation of the log‑transformed MD measure. Pairs of rows marked with the same symbols indicate that these variables were included within the same model.
In the models examining MD measurements alone, all levels of MD were significantly associated with an increased risk of CBC, with an OR of 1.26 (95% CI: 1.01, 1.60; P = 0.047; AUC = 0.638) for Cumulus, 1.33 (95% CI: 1.05, 1.70; P = 0.019; AUC = 0.645) for Altocumulus, and 1.38 (95% CI: 1.09, 1.76; P = 0.009; AUC = 0.646) for Cirrocumulus.
In contrast, the model assessing PRS_MD alone showed no significant association with CBC risk (OR: 1.44, 95% CI: 0.47, 4.48; P = 0.527; AUC = 0.554).
When PRS_MD was included alongside MD measurements, the OR for MD remained stable and unchanged across all models, while PRS_MD continued to show no significant association with CBC risk. However, the addition of PRS_MD resulted in a small increase in AUC across all MD models (e.g., from 0.638 to 0.652 for Cumulus), indicating a modest improvement in model discrimination despite the non‑significant PRS_MD coefficients. To assess whether the lack of independent predictive value for PRS_MD reflected redundancy with MD, we evaluated collinearity between PRS_MD and each MD measure. Pearson correlation coefficients were all < 0.10, and variance inflation factors were < 1.5, indicating low collinearity and suggesting that the minimal contribution of PRS_MD is not due to overlap with measured MD.
We evaluated whether PRS_MD modified the association between MD and CBC risk by including an interaction term (MD × PRS_MD) in multivariable logistic regression models. Interaction terms were not statistically significant for Cumulus (estimate = 0.61, P = 0.303), Altocumulus (estimate = 0.82, P = 0.180), or Cirrocumulus (estimate = 0.90, P = 0.133). These results indicate no evidence that inherited predisposition to MD, as captured by PRS_MD, modifies the effect of measured MD on CBC risk.
Full models can be found in Supplementary Tables S4-S6.

Discussion
In this study, we evaluated the potential role of a PRS constructed from genetic variants associated with MD measurements in predicting different levels of MD as well as CBC risk. Although the PRS was associated with some levels of MD, it did not demonstrate a predictive role for CBC risk. As we previously showed, MD remains a strong predictor of CBC risk, with Cirrocumulus emerging as the most statistically significant MD measure, aligning with our previous analysis22. Our PRS_MD, derived from prior GWAS of MD18 was associated with our density measures. Although adding PRS_MD to models that included MD produced small increases in AUC, these gains were modest and not statistically significant, and PRS_MD was not independently associated with CBC risk. These findings highlight that the current MD PRS does not yet capture sufficient genetic variation to provide information beyond what is already reflected in measured MD.
MD is a well-established risk factor for first primary breast cancer, and its polygenic nature has been extensively studied7,8. MD is influenced by multiple genetic variants, each contributing a small effect to the overall phenotype9–11. While these genetic determinants have been explored in the context of first primary breast cancer, their relevance to CBC risk remains unknown.
Most studies investigating PRS for breast cancer risk have focused on first primary breast cancer23,24, using PRS derived from SNPs associated with breast cancer itself rather than with MD. In this context, MD is typically included as a continuous covariate rather than being used to construct a trait-specific PRS. For instance, Vachon et al. (2019) examined the joint effects of a 77-SNP PRS for breast cancer and quantitative MD measures on breast cancer risk9. Although they demonstrated that both adjusted percent density and absolute dense area added independent predictive value, the study population was limited to women without a prior breast cancer diagnosis, and the study did not address CBC risk.
Likewise, a study within the Nurses’ Health Study and Nurses’ Health Study II integrated a 67-SNP PRS for breast cancer into risk models and found that incorporating this PRS, along with quantitative MD measures and endogenous hormone levels, significantly improved breast cancer risk prediction10. While this supports the value of integrating genetic and imaging-based risk factors, its applicability to CBC risk prediction remains unknown.
To our knowledge, no previous study has evaluated whether a PRS derived from MD-associated SNPs contributes to risk prediction for CBC. Given the unique biological and clinical features of CBC, further research is warranted to assess whether incorporating such a PRS alongside MD could improve individualized risk stratification among breast cancer survivors at risk of developing contralateral breast cancer. Our study directly addresses this gap by testing whether inherited predisposition to MD provides additional predictive information beyond measured MD itself.
In our study, the limited contribution of the PRS_MD to CBC risk prediction could reflect the current state of genetic discovery for MD. The PRS_MD was constructed using 64 genome-wide significant variants from previous GWAS, spanning 37 independent loci18,19. While these variants are associated with MD, they collectively explain only a small fraction (~ 2%) of the variance in MD measurements. This low explanatory power limits the ability of PRS_MD to significantly enhance CBC risk prediction models. This is consistent with the expectation that a PRS explaining little variance in the underlying phenotype is unlikely to meaningfully improve prediction of downstream disease outcomes.
Further, the genetic variants included in PRS_MD were identified primarily in association with MD measured at the conventional Cumulus intensity threshold. These variants may not fully capture the genetic architecture underlying highest-intensity MD measures, such as Cirrocumulus, which may explain the PRS’s reduced performance for this measure. Indeed, our results showed that PRS_MD had a significant association with both Cumulus and Altocumulus densities, but a weaker or no association with Cirrocumulus. This suggests that the genetic determinants of the highest-intensity MD regions may differ from those captured in current GWAS, and future MD-specific GWAS incorporating higher-intensity thresholds may yield more informative PRSs.
We also evaluated whether the MD PRS modified the association between MD and CBC risk. Exploratory interaction analyses did not reveal statistically significant interactions, suggesting that genetic predisposition to MD does not amplify or attenuate the effect of measured MD on CBC risk.
To further contextualize the null findings, we quantified collinearity between PRS_MD and each MD measure. Correlations and variance inflation factors were low, indicating that the lack of independent predictive value for PRS_MD is not due to redundancy with measured MD but rather reflects the limited variance explained by the current PRS.
Evaluation of model discrimination using AUC showed that adding PRS_MD produced small but consistent increases in AUC across all MD models, despite PRS_MD itself not being statistically significant. This suggests that PRS_MD contributes a modest amount of additional information, although not enough to meaningfully improve overall predictive performance in its current form.
The distinct genetic determinants of this highest-intensity threshold warrant further investigation to construct tailored PRS models that better capture its unique characteristics. Although including PRS_MD in the Cirrocumulus model slightly increased the R-squared value, the adjusted R-squared decreased, suggesting that the added complexity from PRS_MD did not meaningfully improve model fit. This discrepancy underscores the possibility that the current PRS_MD may not capture the genetic architecture relevant to the most informative MD measure for CBC risk. Larger cohorts will be crucial to improving the precision and generalizability of findings in future studies.
An additional consideration is that MD was measured primarily from mammograms obtained before the first breast cancer diagnosis. While this design avoids treatment-related changes in MD, it may not fully reflect the density profile at the time CBC develops. MD is a dynamic phenotype influenced by aging, menopausal transition, and treatment, and future studies incorporating longitudinal MD trajectories may provide a more complete understanding of how MD genetics and MD change jointly influence CBC risk.
Our study highlights the potential role of a PRS derived from MD-associated variants in predicting differing aspects of MD. However, this PRS did not improve CBC risk prediction. While MD remains a strong predictor of CBC risk, the limited predictive value of PRS_MD may be due to the small variance explained by known genetic variants and differences in MD measurement methods. Larger studies with refined PRS models are needed to assess its utility in CBC risk prediction.
Based on our findings, the current PRS for MD explains only a small fraction of the variance in MD, limiting its predictive power. However, with future advancements in the genetic discovery of MD and the development of stronger PRS, PRS for MD could serve as a surrogate marker for risk assessment in situations where actual MD measurements are unavailable. This could be particularly useful for younger individuals, populations with limited access to mammography, or large-scale genetic studies. As MD-specific GWAS expand and PRSs capture a greater proportion of MD variance, the potential for MD PRSs to contribute to CBC risk prediction, either independently or in combination with measured MD, may increase. Further research is needed to improve the accuracy and applicability of PRS for MD in clinical risk models.

Methods

Study participants
The WECARE Study is an international, population-based case-control study comparing women with asynchronous CBC to those with unilateral breast cancer (UBC). Eligible participants were identified through population-based cancer registries at each study center. US centers recruited participants from registries contributing to the National Cancer Institute’s Surveillance, Epidemiology, and End Results Program. The participating US centers included Seattle, Washington; Iowa; and California. Additionally, the population-based cancer registry in Ontario, Canada was included.
The WECARE Study consists of two phases, which have been described previously22,25. MD measurements were only collected from a subset of the WECARE Study participants5. In the present analysis, we restricted the sample to women of European ancestry because the available PRS for MD was developed exclusively in European‑ancestry GWAS populations and applying it to other ancestry groups would reduce accuracy and introduce bias. In addition, the subset of WECARE participants with both MD and genotype data was predominantly of European ancestry, limiting statistical power for analyses in other groups.
Participants were diagnosed at least one year after their first primary local or regional-stage invasive breast cancer before age 55 years, between 1990 and 2008. All participants provided written informed consent, and the study received ethical approval from the institutional review boards at the contributing study centers. All methods were performed in accordance with the relevant guidelines and regulations, including the Declaration of Helsinki and its later amendments.

Data and mammogram collection
The WECARE Study participants were interviewed via telephone using a structured questionnaire covering breast cancer risk factors, including demographics, medical history, and lifestyle factors. Information on mammograms conducted before and after the first breast cancer diagnosis was also collected in the second phase of the study.
Medical records provided details on treatment (chemotherapy, hormonal therapy, radiation therapy) and tumor characteristics (e.g., estrogen receptor status). Each study center aimed to obtain two mammograms of the unaffected contralateral breast: one around the time of diagnosis and another 6 to 18 months post-diagnosis, focusing only on the cranio-caudal view. For this analysis, we used only mammograms obtained prior to the first breast cancer diagnosis, thereby avoiding treatment-related changes in MD, including reductions associated with tamoxifen or other hormonal therapies.

Mammogram digitization and MD measurement
Mammogram films were digitized at two locations: Seattle for US mammograms and Toronto for those from Ontario, using a Kodak Lumisys Digital Scanner. Three measures of MD were made at the conventional intensity threshold (‘Cumulus’) and at two sequentially higher-intensity thresholds (‘Altocumulus’ and ‘Cirrocumulus’)26 using the CUMULUS software22. These thresholds correspond to increasingly dense fibroglandular tissue, with Cumulus capturing the standard DA, and Altocumulus and Cirrocumulus capturing progressively higher-intensity, more radiopaque regions of the breast. DA was used as the primary MD phenotype because it aligns with the MD traits evaluated in the GWAS from which the PRS was derived. In this study, the three measures of MD were analyzed separately as outcomes of interest.

Genotyping, quality control, and imputation of non-genotyped variants
Participants from the WECARE Study were genotyped using DNA extracted from saliva samples. Genotyping was performed with the Axiom Precision Medicine Research Array (PMRA)27. Quality control included filtering single nucleotide polymorphisms (SNPs) and samples with < 95% call rate and Hardy-Weinberg Equilibrium (HWE) P < 10⁻⁷. Imputation was performed using the TOPMed server28, excluding SNPs with imputation R² < 0.3 or minor allele frequency (MAF) < 0.05, and 5.95 million SNPs were retained. Ancestry was assessed with PLINK v2.029, and principal components were calculated separately for European-ancestry women using EIGENSTRAT30.
For this project, we included 399 European-ancestry participants from the WECARE Study who had at least one pre-diagnosis mammogram available for analysis with completed CUMULUS assessments and genotype data available.

Estimating polygenic risk score (PRS)
We utilized data from two recent GWAS of MD, derived using CUMULUS software to measure MD. The first study by Sieh et al. (2020) was a meta-analysis involving 24,192 women from the Genetic Epidemiology Research on Adult Health and Aging Study18. The second study by Chen et al. (2022) was a meta-analysis of 30 studies, comprising genotype-level or summary statistic data from 27,900 participants19. These studies did not incorporate the higher intensity threshold measurements (Altocumulus or Cirrocumulus), and thus the PRS constructed for this study is based solely on variants associated with standard Cumulus-defined mammographic density.
A weighted PRS was constructed comprised of 64 previously identified genome-wide significant SNPs (P-value < 5e-08) associated with mammographic phenotype, specifically DA or PDA in the whole breast, identified from the two GWAS18,19. All GWAS effect sizes were standardized (per SD units) in the original studies, making them directly comparable across DA and PDA. For overlapping variants between studies or between DA and PDA within a study, the effect size of the most significant variant was included. When a SNP was unavailable, proxies with R² > 0.8 were identified using LDproxy31. To ensure independence among SNPs included in the PRS, we applied linkage disequilibrium (LD) pruning and removed variants with r² > 0.01. The SNP effect sizes from these GWAS were used to build the PRS.

Statistical models
In the initial phase of our analysis, we assessed the relationship between MD measurements (Cumulus, Altocumulus, and Cirrocumulus) and the PRS using linear regression models incorporating several potential confounders. Each MD measurement was first logarithmically transformed to normalize the data distribution. Subsequently, these values were standardized; each unit increase corresponds to one standard deviation (SD) in the transformed distribution. This standardization facilitated the interpretation of our effect estimates, which are presented as the natural logarithm of the odds ratio (OR) per adjusted SD of the log-transformed variable32. Our models were adjusted for a set of covariates including age and menopausal status at time of mammogram; body mass index (BMI) at time of interview, estrogen receptor status of the first primary breast cancer; treatments received for the first breast cancer (chemotherapy, hormonal therapy, and radiation therapy); age at menarche; and the study center and the year of the mammogram. To quantify the proportion of variance in MD measurements explained by PRS, we calculated both the R-squared and adjusted R-squared values for each model. These metrics provided insights into the explanatory power of PRS, especially after its inclusion in the regression models.
In the second phase of our investigation, we examined whether PRS_MD was an independent predictor of CBC risk beyond MD measurement. Previous findings from the WECARE Study highlighted that higher-intensity MD measurements, specifically those classified as ‘Cirrocumulus’, were statistically significant predictors of subsequent CBC and superior to MD measured by conventional Cumulus22. Building upon these insights, we aimed to determine whether PRS_MD not only serves as an independent predictor of CBC but also enhances the predictive accuracy of existing models when combined with MD measurements.
To evaluate the potential value of adding PRS_MD to the model, we utilized logistic regression models, adjusting for the same confounders as described above. To account for population stratification and its potential confounding effects, the top five principal components were incorporated into all models at the point of introducing PRS_MD. We additionally estimated the area under the receiver operating characteristic curve (AUC) for each model to assess discriminative performance. To evaluate whether the association between MD and CBC risk differed by levels of PRS_MD, we included an interaction term (MD × PRS_MD) in multivariable logistic regression models. Interaction models were adjusted for the same covariates as the main analyses and incorporated the top five principal components to account for population stratification. Collinearity between MD measurements and PRS_MD was assessed using Pearson correlation coefficients and variance inflation factors.

Supplementary Information
Below is the link to the electronic supplementary material.