Introduction

Women with pathogenic or likely pathogenic variants (PV) in BRCA1 or BRCA2 are at increased risk of developing breast, ovarian and other cancers, and may benefit from risk-management strategies such as risk-reducing medication, risk-reducing surgery, or intensive surveillance1. However, the breast cancer risk in BRCA1 or BRCA2 PV carriers varies depending on several factors, including genetics, lifestyle and reproductive risk factors, and also depends on the cancer incidence in the population2,3,4,5,6. There is a need for accurate risk stratification methods to empower women in making well-informed decisions7.

Polygenic risk scores (PRS), combining the effects of multiple disease-associated single nucleotide variations (SNVs), have emerged as powerful tools for breast cancer risk stratification in European populations, both in the general population and BRCA1 or BRCA2 PV carriers3,4,8. Notably, a PRS developed in the European population using 313 SNVs (PRS313) has shown predictive value for breast cancer risk in Asian populations, albeit with lower discrimination compared to women of European descent9. Subsequently, the predictive performance of this European-based PRS was further enhanced by incorporating additional SNVs associated in Asian studies, resulting in a PRS with 333 SNVs (PRS333)10. Additionally, Ho et al.10 generated a PRS using a trans-ancestry Bayesian polygenic prediction approach implemented in PRS-CSx11, resulting in a PRS that includes approximately one million SNVs (PRSGW) and demonstrates the highest predictive accuracy compared to the other two PRS in predicting breast cancer risk in the general Asian populations.

In this study, we aimed to evaluate if these PRS can predict the risk of breast cancer in BRCA1 and BRCA2 PV carriers of Asian ancestry.

Results

The study cohort consists of 604 BRCA1 (390 affected with breast cancer, 96 with ovarian cancer, 118 unaffected with either cancer) and 785 BRCA2 (552 affected with breast cancer, 46 with ovarian cancer, 187 unaffected) PV carriers recruited from 4 different countries (Supplementary Table S1). The mean age of diagnosis for BRCA1 PV carriers was 40.9 (SD = 10.2 and for BRCA2 PV carriers was 44.3 (SD = 10.2), while the mean censoring age for unaffected individuals for BRCA1 PV carriers was 46.5 (SD = 13.9) and for BRCA2 carriers was 43.4 (SD = 13.8) (Table 1).

Table 1 Description of study cohort
Full size table

PRS association with breast cancer risk

Five PRS, which were previously demonstrated to be predictive for breast cancer risk in the general Asian population,n were selected for analysis: three based on PRS developed through analyses of studies in the Breast Cancer Association Consortium in European populations8, and two based on trans-ancestry analyses9,10 (Supplementary Table S2). We evaluated the association of each PRS with breast cancer risk. The strongest association, in both BRCA1 and BRCA2 PV carriers, was with PRSGW (HR = 1.47, 95% CI = 1.10–1.95, p = 0.0089 and HR = 1.43, 95% CI = 1.04–1.95, p = 0.0255, respectively) (Table 2). The estimated HRs for the other PRS were greater than 1 but smaller and not statistically significantly different from 1. Sensitivity analysis showed that the hazard ratios for PRSGW were slightly lower and not significant for BRCA1 (HR = 1.31, 95% CI = 0.94–1.83, p = 0.1119) and similar for BRCA2 PV carriers (HR = 1.44, 95% CI = 1.01–2.07, p = 0.0457) when the weights were calculated based on respective country incidence rates rather than using average incidence rates. (Supplementary Table S3).

Table 2 Association of PRSs with BC risk in BRCA carriers
Full size table

We further evaluated the association between PRSGW and breast cancer risk when the PRS was treated as a categorical variable. Compared to women in the middle quintile (40–60%), the estimated HRs for developing breast cancer for women in quintiles 1 and 5 were 0.89 (95% CI 0.52–1.52) and 1.57 (0.99–2.49), respectively, for BRCA1, and 0.44 (0.24–0.80) and 1.09 (0.67–1.80) (Table 3) for BRCA2 PV carriers. The estimated HRs by PRS percentile did not differ from those predicted under a theoretical polygenic model in which the log HR depends linearly on the PRS: all predicted HRs fell within the confidence intervals of the observed HRs.

Table 3 Proportion of samples in percentile categories of PRSGW and their associations with breast cancer risk
Full size table

Predicted absolute risk by PRS percentile

We used the hazard ratio estimates and the average breast cancer incidence across the four countries included in this study to compute age-specific absolute cumulative breast cancer risks for PV carriers by PRS percentiles according to PRSGW (Fig. 1). BRCA1 PV carriers at the 5th percentile of the PRS distribution had an estimated risk of 17% of developing breast cancer by age 50 years and a 38% risk by age 80 years. In contrast, the BRCA1 PV carriers at the 95th percentile of the PRS distribution had a 45% breast cancer risk by age 50 years and 81% by age 80 years. BRCA2 carriers at the 5th percentile of the PRS distribution had a risk of 10% of developing breast cancer by age 50 years and a 28% risk by age 80 years. In contrast, BRCA2 carriers at the 95th percentile of the PRS distribution had a 30% breast cancer risk by age 50 years and 65% by age 80 years.

Fig. 1: Age-specific absolute cumulative breast cancer risks for PV carriers by PRS percentiles according to PRSGW.
figure 1

a BRCA1 PV carriers and (b) BRCA2 PV carriers.

Full size image

Discussion

In this study, we evaluated the association of PRS, previously validated in the Asian general population, with the risk of breast cancer among carriers of BRCA1 and BRCA2 PVs of Asian ancestry. While all estimated HRs were above one 1 for all PRS, the association was statistically significant only for the PRSGW, constructed using approximately 1 million variants. The magnitude of the association was markedly higher than for the other PRS in both BRCA1 (1.47 versus 1.12–1.16) and BRCA2 (1.43 versus 1.04–1.10) PV carriers.

This is qualitatively consistent with our previous observations in the general population, which showed that PRSGW (HR = 1.62, 95%CI:1.46–1.80) outperformed PRS333 (HR=1.53, 95%CI:1.37–1.71) and PRSOVERALL (HR = 1.46, 95%CI:1.34–1.60) in prospective cohorts10, although the difference was much less marked. PRSGW was developed through integrating GWAS summary statistics from multiple populations and leveraging linkage disequilibrium diversity across discovery samples11. This enables more accurate effect size estimation and hence improves the predictive accuracy of PRS in the target population.

We found that the estimated effect sizes for all of the PRS on cancer risks in BRCA1 and BRCA2 PV carriers were lower than those previously observed in the general population: for PRSGW, the corresponding HR in Asian prospective cohorts was 1.62 (95%CI:1.46-1.80). Again, attenuation is qualitatively consistent with what was observed in European populations (e.g., for the 313SNV PRS, the HR was estimated to be 1.61 in the general population compared to 1.20 and 1.31 in BRCA1 and BRCA2 PV carriers4,8).

In the European studies, the PRS adapted for ER-negative breast cancer (PRSER-) was shown to be more predictive than PRSOVERALL for breast cancer risk in BRCA1 PV carriers, consistent with the known strong association of BRCA1 PVs with ER-negative (specifically triple negative) breast cancer, while PRSOVERALL was more predictive for BRCA2 PVs. PRSOVERALL was clearly weaker in BRCA1 than BRCA2 PVs. This difference was not apparent in this study, possibly due to chance, given the wide confidence limits.

Previous research by Kuchenbäcker et al.3, indicating improved performance of an 88-SNV PRS in the European population after incorporating BRCA-specific SNVs identified from BRCA1 and BRCA2 specific genome-wide association studies (GWAS) compared to 77-SNV PRS derived from the general population. This suggests including variants that are associated with cancer risk in BRCA1 and BRCA2 PV carriers might potentially further improve the predictive accuracy of PRS. However, SNVs for Asian BRCA PV carriers remain elusive as current GWAS in BRCA PV carriers lack representation from Asian populations.

We showed that PRSGW can achieve a useful level of risk stratification in Asian BRCA1 and BRCA2 PV carriers, where the cumulative risk was substantially lower for carriers in the lowest PRS percentile compared to the highest PRS percentile (Fig. 1). A previous study has shown that the absolute breast cancer risk of Asian PV carriers varies depending on the underlying population-specific cancer incidence6. Asian carriers residing in countries with significantly lower population cancer incidences are expected to have markedly lower absolute cancer risks compared to European ancestry carriers. While risk stratification by PRS may not alter screening recommendations, it can refine risk assessment. For instance, carriers identified as having lower risk may consider delaying prophylactic surgeries, such as mastectomy or oophorectomy, thus balancing the benefits and potential harms of such interventions. Implementation of PRS comprising such an expansive SNV set can be difficult in practice. However, given the significant findings, PRSGW can be an alternative until larger GWAS of Asian ancestry BRCA1 and BRCA2 PV carriers become available for the development of BRCA-specific PRS.

One limitation of our study is that although this is the largest available dataset for Asian BRCA1 and BRCA2 PV carriers, our study may lack the power to detect associations with PRS of marginal magnitude. Moreover, the confidence limits associated with the HR estimates were wide. Sensitivity analyses utilizing weights derived from country-specific cancer incidence rates yielded generally comparable HRs to those derived from average cancer incidence rates. However, the statistical significance was reduced in the sensitivity analysis, particularly among BRCA1 PV carriers. This attenuation may be due to additional heterogeneity introduced when applying country-specific incidence rates to stratified age groups, particularly where sample sizes within certain age-country strata were small (Supplementary Table S4). These findings highlight the challenges in achieving robust statistical power in subgroup analyses and underscore the need for larger, well-powered studies in diverse populations.

In summary, the results demonstrate the potential utility of PRSGW in predicting the risk of breast cancer for Asian carriers of both BRCA1 and BRCA2 PVs. Incorporating this polygenic risk score into risk prediction models for PV carriers, alongside other risk modifiers, may be crucial for refining population-specific cancer risk assessments, especially for Asian carriers with a lower risk of breast cancer.

Methods

Study population

Eligible study subjects included in the Consortium of Investigators of Modifiers of BRCA1/2 (CIMBA) were self-reported Asian female carriers of a pathogenic or likely pathogenic variant (PV) in either BRCA1 or BRCA2 who were age 18 years or older. The germline mutations were classified as pathogenic or likely pathogenic if they resulted in a truncated protein or have been previously reported as disease-associated by ClinVar (https://www.ncbi.nlm.nih.gov/clinvar/) or ENIGMA BRCA1/2 expert panel guidelines (https://enigmaconsortium.org/) based on ACMG/AMP Guidelines. Carriers with variants of uncertain significance were excluded. Carriers were recruited from seven study centres in four countries, as part of either population or hospital-based case-control studies, or through genetics clinics (Supplementary Table S1). All research reported here was performed in accordance with the Declaration of Helsinki and each of the study centres recruited carriers under protocols approved by local ethics review boards. The Malaysian Breast Cancer Genetic Study (MyBrCa) was approved by the Independent Ethics Committee, Ramsay Sime Darby Health Care (reference no: 201109.4 and 201208.1), and the Medical Ethics Committee, University Malaya Medical Centre (reference no: 842.9). SGBCC was approved by the National Healthcare Group Institutional Review Board (NHG DSRB Ref: 2009/00501, approval date 16 December 2009) and the SingHealth Duke-NUS Institutional Review Board (CIRB Ref: 2019/2246, approval date 29 October 2010). Each study listed was approved by the local institutional ethics committees and review boards. Written informed consent was obtained from all subjects. Subjects without genotype data were excluded from the analyses, leaving 604 BRCA1 (390 affected with breast cancer, 96 with ovarian cancer, 118 unaffected with either cancer) and 785 BRCA2 (552 affected with breast cancer, 46 with ovarian cancer, 187 unaffected) PV carriers in this study. Blood samples were genotyped with the iCOGS array or Oncoarray, which provides genome-wide genotyping12,13. Standard quality control processes applied to the genotype data have been described in detail elsewhere14,15: these, which included assessment of the SNV call rate, allele frequency, genotyping intensity clustering metrics, Hardy-Weinberg equilibrium and SNV concordance in duplicate samples. Genotypes for variants not on the arrays were estimated using two-stage imputation, using SHAPEIT and IMPUTE2, with the 1000 Genomes Project (Phase 3) samples as the reference panel.

Polygenic Risk Score

PRS were computed using the standard formula:

$${PRS}=\,{\beta }_{1}{x}_{1}+{\beta }_{2}{x}_{2}+...+{\beta }_{k}{x}_{k}+\ldots +{\beta }_{n}{x}_{n}$$
(1)

where xk is the dosage of risk allele (0-2) for SNV k and βk is the corresponding log odds ratio for SNV k.

The list of SNVs and their corresponding log odds ratios is in accordance with those reported in previous publications. PRSOVERALL was based on the 313 SNV PRS developed by Mavaddat et al. For these analyses, the PRS was restricted to the 287 SNVs with imputation accuracy >0.9 in the Asian studies (that is, the original weights were used for the 287 SNVs, but weights for the remaining 26 SNVs were set to zero)9. PRSER+ and PRSER- are modified versions of PRSOVERALL in which the weights from Mavaddat et al optimised for the prediction of ER+ and ER- negative, were used. The trans-ancestry PRS, PRS333, was a weighted average of the European-based PRSOVERALL and 46 SNV PRS derived from GWAS in Asian populations, as given in Ho et al.10. Thus, PRS333 was derived using:

$${{\rm{PRS}}}_{333}={{\rm{\alpha }}}_{1}{{\rm{PRS}}}_{{\rm{ASN}}}+{{\rm{\alpha }}}_{2}{{\rm{PRS}}}_{{\rm{OVERALL}}}+{{\rm{\alpha }}}_{0},$$
(2)

where α1 = 0.14893, α2 = 0.35354, and α0 = -0.05224, and PRSASN was a 46 SNV PRS. The final PRS, PRSGW, was derived as a weighted average of European and Asian-specific PRS, generated using a Bayesian polygenic prediction model in PRS-CSx, thus:

$${{\rm{PRS}}}_{{\rm{GW}}}={{\rm{\alpha }}}_{1}{{\rm{PRS}}}_{{\rm{GW\_ASN}}}+{{\rm{\alpha }}}_{2}{{\rm{PRS}}}_{{\rm{GW\_EUR}}}+{{\rm{\alpha }}}_{0},$$
(3)

where α1 = 0.16856, α2 = 0.38484, and α0 = 0.54881. Lists of SNVs and the corresponding weights as describe in the original articles.

To facilitate a direct comparison of the performance of each PRS, we standardized the PRS to the standard deviation (SD) of the PRS in the validation set of control subjects previously reported9,10.

Statistical analysis

The association between each PRS and the incidence of breast cancer was evaluated in a survival analysis framework. Individuals were considered at risk from birth and censored at the age of the first breast or ovarian cancer diagnosis, age at bilateral prophylactic mastectomy, or the age at last follow-up. There were two women in the study with censoring age > 80 (both with age of last follow-up at age 81). PV carriers censored at ovarian cancer diagnosis were considered unaffected for the breast cancer analysis. To account for the oversampling of affected BRCA1 and BRCA2 PV carriers, the association of each PRS with breast cancer risk was analysed using a weighted cohort Cox regression with time to breast cancer diagnosis as the outcome16. This method involves assigning different weights to affected and unaffected individuals, which are age- and gene-specific, so that the weighted observed incidence rate aligned with externally derived incidence rates for carriers. The country-specific breast cancer incidence rates for BRCA1 and BRCA2 PV carriers were estimated using country-specific population breast cancer incidence, the reported log relative risks and the method described in Ho et al.6, where the log relative risks were assumed to be the same across all countries. The weights for non-random sampling adjustment were calculated based on the average breast cancer incidence rates in BRCA1 or BRCA2 PV carriers across all countries. The estimated BRCA1 and BRCA2 PV carrier breast cancer incidence rates and the corresponding weights are provided in Supplementary Table S5 and Supplementary Table S6, respectively.

PRS was treated as either a continuous or a categorical variable in the model. The first 4 ancestry principal components (PCs) and birth cohort (in decades) were included as covariates. The robust variance approach was used to account for related individuals in the study by clustering on family membership. All models were fitted separately in BRCA1 and BRCA2 PV carriers. When used as a categorical predictor, the PRS was grouped into quintiles based on the PRS distribution in unaffected PV carriers. The middle group (40–60%) was used as the reference category. The observed HRs by PRS percentiles were compared with the theoretical HR predictions under a multiplicative polygenic model of inheritance17. The weighted cohort analysis was carried out in R “survival” library command coxph(model,robust = TRUE,weights = w) where w represents the age-specific weights.

The age-specific absolute risks of developing breast cancer in each PRS percentile were calculated using the following formula described in Barnes et al.4:

$${\mathrm{AR}}_{{\rm{g}}}\left(t\right)=\mathop{\sum }\limits_{u=0}^{t}{\lambda }_{{\rm{g}}}\left(u\right)\cdot {{\rm{S}}}_{{\rm{g}}}(u)$$
(4)

where \({\lambda }_{{\rm{g}}}\)(u) = \({\lambda }_{0}\left(u\right)\exp {(\beta }_{g})\) is the estimated breast cancer incidence associated with PRS at age u, with \({\lambda }_{0}\left(u\right)\) representing the baseline incidence and \({\beta }_{g}\) the corresponding log hazard ratio of association with breast cancer risk for PV carriers in PRS category g relative to the reference category. Here, Sg(u) is the probability of being breast cancer free at age u. The PRS-specific breast cancer incidences, \({\lambda }_{{\rm{g}}}\)(u), were calculated iteratively by assuming that the average age-specific breast cancer incidence over all PRS percentiles agreed with the estimated average BRCA1 or BRCA2 PV carrier breast cancer incidence.

All statistical analyses were conducted using R v.3.6.3.