Using Polygenic Risk Scores to Refine Risk and Outcomes in High-grade Serous Ovarian Cancer in Australia

Abstract

Background:

Ovarian cancer is characterized by high mortality and lacks effective screening, making prevention critical. A polygenic risk score (PRS), which aggregates the effects of multiple common alleles, may capture a proportion of currently unexplained genetic risk. Although PRSs have been evaluated for risk prediction, their association with treatment response and survival remains unclear. This study assessed the utility of a PRS for predicting high-grade serous ovarian cancer (HGSOC) risk in an Australian population and its association with chemotherapy response and outcomes.

Methods:

PRSs were calculated for 1,097 cases with HGSOC and 812 controls using data from Australian research programs. Associations among PRS, ovarian cancer risk, chemotherapy response, and survival were analyzed.

Results:

Each standard-deviation increase in PRS was associated with a 40% increase in HGSOC risk [odds ratio (OR), 1.40; P < 0.001]. Women in the top 1% of the PRS distribution had a lifetime ovarian cancer risk approaching 3%. Higher PRS values showed a trend toward poorer outcomes; however, these associations were not consistent across analyses.

Conclusions:

PRSs were not clearly associated with chemotherapy response or survival but represent a significant risk factor for the development of HGSOC. Incorporating PRS into clinical models may improve risk stratification and support targeted prevention.

Impact:

As the first study to evaluate how PRS relates to both HGSOC risk and chemotherapy response and outcomes, we show that PRSs are unlikely to serve as therapeutic biomarkers but support their use for enhanced risk-stratified prevention.

Introduction

Every year, approximately 240,000 women around the world, including 1,800 Australian women, are diagnosed with ovarian cancer, the majority of whom present with advanced stage disease (1–3). Ovarian cancer is a heterogeneous disease comprising several distinct malignancies occurring at the same anatomic site (4). Epithelial ovarian cancer (EOC), including serous, mucinous, endometrioid, and clear-cell subtypes, is the most common type of ovarian cancer, with high-grade serous ovarian cancer (HGSOC) being the predominant and most aggressive form, originating in the fallopian tube (5, 6).

Despite advances in the understanding of HGSOC biology and treatment, the survival rate for women with HGSOC is still less than 50% five years after diagnosis (3, 7). Plagued by a lack of effective screening (8) and the almost ubiquitous development of chemotherapy resistance, prevention of HGSOC remains the most effective mechanism to reduce the morbidity and mortality of this disease. When performed at the recommended age, risk-reducing salpingo-oophorectomy (RRSO) reduces ovarian cancer risk by more than 95% and improves all-cause mortality among women with a pathogenic variant (PV) in BRCA1 or BRCA2 (9–12). With data supporting efficacy and cost-effectiveness, RRSO is accepted internationally as a cancer prevention strategy for women with a lifetime risk of ovarian cancer of at least 5% (13, 14).

There are now 11 proven ovarian cancer susceptibility genes (BRCA1, BRCA2, BRIP1, RAD51C, RAD51D, PALB2, MLH1, PMS2, MSH6, MSH2, and EPCAM; ref. 15). Testing for germline PVs in these genes is performed routinely in the clinical setting, and the option of RRSO made available to family members with a PV, who are yet to be personally affected by ovarian cancer, as a cancer prevention intervention. However, it is now recognized that genetic heritability extends beyond single-gene variants (16, 17) and is likely influenced by the aggregate effect of single-nucleotide polymorphisms (SNP), of which 27 discovered through genome-wide association studies (GWAS) have been found to account for around 6% of ovarian cancer heritability (16, 18, 19). Individually, these SNPs are associated with a minor increase or decrease in an individual’s personal risk of ovarian cancer. A polygenic risk score (PRS) quantifies the risk associated with the inheritance of multiple risk-associated SNPs and consequently an individual's propensity to malignancy (20, 21). In a small number of validation studies, an increased PRS has been associated with an increased risk of developing ovarian cancer (22, 23). However, ovarian cancer incidence has significant geographic variability. Whether PRS performance is the same in an Australian population with a lower background of ovarian cancer incidence [13 per 100,000 (3) vs. 25 per 100,000 (24) in the United Kingdom] is unknown.

The biological effects of variation in polygenic risk remain incompletely understood. More than 90% of GWAS variants fall in noncoding regions of the genome, with an enrichment in DNA regulatory elements, which suggests that SNPs may act through regulation of gene expression (25–27). For example, variants occurring within promoters or enhancers have the potential to alter epigenetic states or directly affect gene expression through disruption of transcription factor binding (28). Epigenetic alterations, such as aberrant promoter methylation, have been reported to be associated with chemotherapy response (29, 30). A previous study analyzed high-resolution H3K27ac-HiChIP chromatin contact maps generated from ovarian cancer cell lines and revealed that 162 target genes of GWAS risk loci were enriched in cancer-related pathways, including the homologous recombination pathway (31). It is well established that harboring a germline PV in BRCA1/2 is associated with improved ovarian cancer survival in the first 5 years (32). A prognostic benefit is also seen in patients with homologous recombination deficiency (HRD) without a PV in BRCA1/2 (33, 34), which may, in part, be explained by improved platinum sensitivity (35). Thus, if ovarian cancer–associated SNPs, particularly those incorporated into a PRS, are enriched in HRD-related pathways, they may have implications not only for cancer risk but also for platinum response and patient survival.

The aim of this study was to evaluate the utility of a PRS for identifying Australian women at an increased risk of HGSOC, to inform targeted cancer prevention strategies. Specifically, we assessed the performance of the PRS within an Australian population and examined its association with treatment response and clinical outcomes in women with HGSOC, focusing on whether PRS predicts chemotherapy sensitivity and survival.

Materials and Methods

Ethics approval and consent

This research was reviewed and approved by the local human research ethics committee (HREC) at the Peter MacCallum Cancer Centre (Melbourne; HREC/17/PMCC/89) and was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki. Participants from existing research studies had consented to genetic research through approved studies {16/161 [Australian Ovarian Cancer Study (AOCS)], Peter MacCallum Cancer Centre; P287 [Australian Cancer Study (ACS)] and P1424 [Ovarian Cancer Prognosis And Lifestyle Study (OPAL)], QIMR Berghofer}.

Study population

The study cohort comprised cases and controls (women without ovarian cancer) obtained from three existing research programs: the AOCS (36), the ACS (37), and the OPAL (38). These studies contributed clinical (Supplementary Table S1) and genotyping data, including ovarian cancer susceptibility gene germline status. The AOCS (including ACS) and OPAL recruited women newly diagnosed with ovarian cancer 18 to 79 years of age in hospitals across Australia between 2001 to 2006 and 2012 to 2015, respectively. Controls were frequency-matched to cases by age (5-year groups) and state of residence and were identified from the Australian Electoral Roll. The research programs collectively contributed participant samples to the Ovarian Cancer Association Consortium (OCAC), a multidisciplinary consortium that has generated a large volume of clinical and genotyping data, including SNP profiling, which was used in this study (19, 39, 40). Cases with ovarian cancer were also genotyped through TRACEBACK, a cancer prevention project involving sequencing of germline samples from ovarian cancer cases that had not previously undergone genetic testing (41). Due to the potential for participant overlap between the TRACEBACK and OCAC cohorts, case identification numbers were cross-checked to identify duplicate genotyping records. Eighty-seven cases were found to have undergone genotyping through both studies. Genotype concordance between the two independent platforms was high (94%). To avoid potential bias arising from duplicate data, TRACEBACK-derived genotypes were excluded for overlapping cases.

Clinical data including age, ethnicity, year of diagnosis, histology, stage, treatment, ovarian cancer susceptibility gene germline status, and survival outcomes were collected on all cases (Supplementary Table S1). For controls, only age, ethnicity, and family history were available.

Due to the likelihood that a number of tumors historically identified as high-grade nonmucinous ovarian cancer (HGNMOC) would be considered HGSOC if diagnosed today (42, 43), the study included an analysis of PRS in all HGNMOC cases compared with the subset of HGSOC, allowing an exploration of the influence of histotype.

Risk alleles

Twenty-five common risk alleles (Supplementary Table S2) associated with HGSOC and the corresponding odds ratios (OR) were chosen based on the largest ovarian cancer GWAS conducted at the time of study design (19). Although this GWAS included all EOC subtypes, we selected variants and corresponding effect estimates derived from HGSOC-specific analyses that reached genome-wide significance (P < 5 × 10⁸) for inclusion in this study. Three additional SNPs published by Kar (44) that did not reach genome-wide significance but were regarded by OCAC investigators as likely to be associated with HGSOC (rs17041869, P = 0.00033; rs7937840, P = 0.00015; and rs1469713, P = 0.0171808) were also included. Although there was some variability in the annotation of the alleles, the OR for each risk loci was consistent between the OCAC database and the Phelan GWAS (allowing for rounding errors; Supplementary Table S2; ref. 19).

These 25 variants occur in noncoding regions (intronic or intergenic) and are single-nucleotide changes, with the exception of two variants (rs144962376 and rs8176685) that are small insertions/deletions (indel). For simplicity, all variants are referred to hereafter as SNPs. Following quality control (QC), two variants were excluded; SNP rs8176685 consistently failed QC because of low read counts, and rs6755777 was excluded because of high linkage disequilibrium with rs711830 (R² = 0.991). The remaining variants were evaluated as part of the PRS model. Two additional SNPs (rs2165109 and rs7937840) demonstrated minimal contribution to PRS performance and were therefore not retained in the final model. The final PRS comprised 21 SNPs (Supplementary Table S3).

Genotyping of SNPs in the TRACEBACK cohort was performed using a custom-designed Agilent SureSelect capture panel as previously described (41). Read counts and frequencies for the SNP and indel positions were generated using the “bam-readcount” tool v0.8.0 (45). A minimum read depth of 10 was required for SNP calling, and samples with fewer than 96% of SNP genotypes successfully called were excluded from PRS calculation. Variant allele frequency, referred to here as risk allele frequency (RAF), was calculated as the proportion of reads supporting the risk allele at each position. Genotypes were assigned using the following thresholds:

• RAF of <0.25; homozygous non-risk allele;

• RAF ≥0.25 and ≤0.75; heterozygous risk allele;

• RAF >0.75; homozygous risk allele.

Genotypes were categorized accordingly and recorded for downstream analysis.

Genotyping data for additional cases and controls from the OCAC cohort were generated using the Illumina OncoArray platform, which assays approximately 533,000 variants (19, 39, 40). Genotypes were provided as imputed dosage files representing the expected number of minor alleles. These dosages were converted into integer genotypes (0, 1, 2) using custom R scripts (version 3.6.3). Values with decimal components greater than 0.85 or less than 0.15 were rounded to the nearest integer, whereas intermediate values were set to missing (−9). Genotypes were then translated into allelic categories (heterozygous, homozygous risk, or homozygous non-risk) based on the designated risk allele.

Missing or failed SNP calls were accounted for using an imputed OR, calculated as: imputed OR = LN [(2 * RAF * OR_Hetero (1 − RAF)) + (OR_Homo (RAF²)) + (1 − RAF)²], where LN denotes the natural logarithm.

Defining weight of the risk allele

Histotype-specific ORs were used as weights for each of the chosen SNPs. The β-coefficients obtained from the OCAC consortium were converted to ORs using the exponential function. Log ORs (LOR) were then calculated as follows:

• Heterozygous risk allele = LN(OR);

• Homozygous risk allele = LN(OR²);

• Homozygous non-risk allele = LN (0).

Calculation of the PRS

The raw PRS was calculated for cases and controls by summating the LOR (46) using the equation; PRS = β₁x₁ + β₂x₂ + … β_kx_k… + β_nx_n.

The PRS was then standardized by adjusting for the contribution of each SNP based on the measured allele frequency and standard deviation (SD) of the resulting PRS in the control dataset to provide a Z-score.

For comparison, an expected unadjusted mean based on reported population allele frequencies was calculated. The RAFs were obtained from TOPMed via the Single Nucleotide Polymorphism Database (47), except for rs10962692 and rs144962376, the RAFs of which were obtained from gnomAD and 1000 Genomes, respectively, with the mean calculated as (46)

$$\mathrm{\mu }\mathrm{=}2{\displaystyle \mathrm{\sum }}_k{\mathrm{p}}_k{\mathrm{\beta }}_k$$

where p_k denotes RAF for SNP k and β denotes natural LOR.

Similarly, the variance was calculated as

$${\sigma }^2\mathrm{=}{\displaystyle \mathrm{\sum }}_k{\sigma }_k^2\mathrm{=}\mathrm{\hspace{0.17em}}2{\displaystyle \mathrm{\sum }}_k{p}_k{q}_k{\beta }_k^2$$

where q_k is calculated as 1 − p_k, and the SD is the square root of the variance (46).

Clinical information

Detailed clinical information on cases was collected from the ovarian cancer research programs (Supplementary Table S1). Overall survival (OS) was defined as the interval between histologic diagnosis, or in some circumstances commencement of primary treatment, and death. Progression was determined by clinical assessment, CA-125 levels, and/or imaging results (as per the Response Evaluation Criteria in Solid Tumors guidelines modified for ovarian cancer; ref. 36). Progression-free survival (PFS) was defined as the interval between histologic diagnoses (or in some circumstances commencement of primary treatment) and first progression, death, or last follow-up. Platinum-free interval (PFI) was defined as the time (in days) from last chemotherapy cycle (regardless of drug) to the date of progression for patients treated with platinum. Time until second-line therapy and time until third-line were defined as the time in days from diagnosis to the start of second-line chemotherapy and third-line chemotherapy, respectively.

Statistical analyses

Absolute cumulative risk of ovarian cancer

Absolute cumulative risk of ovarian cancer was estimated by combining age-specific population incidence rates with relative risk (RR) estimates derived from our regression model. Specifically, age-specific ovarian cancer incidence rates for the Australian female population were obtained from the Australian Institute of Health and Welfare (AIHW) cancer data portal (https://www.aihw.gov.au/reports/cancer/cancer-data-in-australia/data). These baseline incidence rates were scaled by RR estimates corresponding to each percentile of the PRS distribution and age. RRs (expressed as ORs) were derived from an age-adjusted ovarian cancer risk model fitted to the study data. The model assumed proportional effects of the PRS across age and did not incorporate birth cohort effects on ovarian cancer incidence. Absolute cumulative risk for a specified age was then calculated by integrating the adjusted age-specific incidence rates across age.

Assessing the impact of PRS on treatment response and outcome

Multivariable Cox proportional hazards models for survival outcome (survival time) and treatment outcome (PFI, second-line therapy, and time until third-line treatment) were performed to determine whether the association between PRS and treatment outcome remained significantly associated with survival after adjusting for both continuous (age) and categorical (residual disease and stage) prognostic variables.

The Cox proportional hazards method investigated the association between the prognostic factors, including PRS for the time-specified outcomes:

1. 10-year OS;

2. 5-year OS;

3. 10-year PFS;

4. 5-year PFS;

5. PFI;

6. Time from diagnosis until second-line treatment;

7. Time from diagnosis until third-line treatment.

In the Cox proportional hazards model, the hazard for an individual $i$ was denoted as follows:

$${\mathrm{\lambda }}_i\left(t\right)\mathrm{=}{\mathrm{\lambda }}_0\left(t\right)e^{X_i\beta }$$

where ${\mathrm{\lambda }}_0$ is the baseline hazard, which is a nonnegative function of time, and $\mathrm{\beta }$ is a $p\mathrm{\times }1$ column vector of coefficients of the prognostic factors. The hazard ratio (HR) for two subjects with fixed covariates ${\mathrm{X}}_{\mathrm{i}}$ and ${\mathrm{X}}_{\mathrm{j}}$ is constant over time as follows (48):

$$\frac{{\mathrm{\lambda }}_i\left(t\right)}{{\mathrm{\lambda }}_j\left(t\right)}\mathrm{=}\mathrm{\hspace{0.17em}}\frac{{\mathrm{\lambda }}_0\left(t\right)e^{X_i\beta }}{{\mathrm{\lambda }}_0\left(t\right)e^{X_j\beta }}\mathrm{=}\frac{e^{X_i\beta }}{e^{X_j\beta }}$$

Each model was then tested for the proportional hazard assumption using Schoenfeld residuals (48). If the proportionality assumption was violated for a categorical prognostic variable, then it was used as a stratification variable in the model. The strata divided the subjects into disjoint groups, where the hazard for an individual $i$ belonging to stratum $k$ is

$${\mathrm{\lambda }}_k\left(t\right)e^{X_i\beta }$$

Similarly, if a continuous variable violated the proportionality assumption, then nonlinear effects were explored using a restricted cubic spline (48–50) to model nonproportional hazards.

The log relative hazards were then computed for the prognostic factors in the model as follows:

$$\ln \left(\frac{\mathrm{\lambda }\left(t\right)}{{\mathrm{\lambda }}_0\left(t\right)}\right)\mathrm{=}{\beta }_1{X}_1\mathrm{+}\mathit{\dots }{\mathrm{+}\beta }_p{X}_p$$

where $\mathrm{\lambda }\left(t\right)$ is the expected hazard at time $t$ and ${\beta }_1$ and ${\beta }_2$ represent the regression coefficients of the prognostic factors.

In some models, as mentioned above, restricted cubic splines (nonlinear components) were included for continuous variables to model the nonproportional hazards. For continuous variables, the log relative hazard of the variable was plotted for a range of its values by holding other variables constant, whereas for categorical variables forest plots were used.

Assessing the association between PRS and HGSOC risk

Univariable and multivariable logistic regression was conducted. This included analysis both including and excluding the carriers of a pathogenic germline variant in an ovarian cancer susceptibility gene (PV). In the univariable logistic regression, ovarian cancer status (case or control) was examined with PRS treated as either a continuous or categorical variable. As a categorical variable, PRS was divided into three groups: low, medium, and high within the control group. In the multivariable logistic regression, both PRS and age were used (age was included as an adjustment factor to account for the matching). The age-related cumulative absolute risk for ovarian cancer according to the PRS percentiles was calculated using coefficients from the multivariable model (excluding PV carriers) and the absolute risks for the Australian population obtained from the AIHW Cancer Data in Australia resource (https://www.aihw.gov.au/reports/cancer/cancer-data-in-australia).

Assessing the impact of PRS and other prognostic factors on survival outcomes

Multivariable Cox proportional hazards models were developed to assess the association of the PRS score, adjusting for age at diagnosis, disease stage, and residual status, with the following endpoints: OS (defined as the interval between diagnosis and death; overall and administrative censoring at 5 and 10 years), PFS (the interval between diagnosis and first progression or last follow-up; overall and administrative censoring at 5 and 10 years), PFI, and time from diagnosis until second-line therapy and time until third-line treatment (Supplementary Table S1). We evaluated potential nonlinear associations of age at diagnosis and PRS (continuous) with the outcome using likelihood ratio tests. The proportional hazards assumption was assessed using the Grambsch–Therneau test (51) based on scaled Schoenfeld residuals and verified through graphical methods (52). Stratified Cox regression models were considered to alleviate any nonproportional issues with categorical variables (48, 52). Final models were internally calibrated and validated using bootstrap sampling (52).

PRS clinical statistics

Differences in categorical variables were assessed using χ² tests. Group differences for continuous variables were examined using the independent t test or Mann–Whitney test, based on the Shapiro–Wilk test for normality. All analyses were performed using GraphPad Prism 8.4 (GraphPad Software), R Statistical Software (v4.1.2; R Core Team, 2021; RRID: SCR 001905), and “rms” R package (v6.2.0; Frank E Harrell Jr, 2021; RRID: SCR_023242). Two-sided P values of <0.05 were considered statistically significant.

Results

Study population

Data from 1,578 women with HGNMOC (626 TRACEBACK and 952 OCAC genotyped) and 812 controls (OCAC genotyped) were included in the analysis (Fig. 1). Most women were less than 70 years (mean age: 61 years in cases and 57 years in controls) and of presumed European ancestry [recorded as “White” (94%, 1,319/1,405 cases; 95%, 772/812 controls); Supplementary Table S4)]. The majority of cases had HGSOC histology (70%, 1,097/1,578) and advanced disease at presentation (74%, 1,161/1,578; Supplementary Table S4). Although neoadjuvant chemotherapy use increased over time, increasing from 13% between 2000 and 2005 to 38% after 2012 (P value <0.0001; χ² test), most cases underwent primary surgery and adjuvant chemotherapy (Supplementary Table S4).

Figure 1.

Overview of the study cohort. Data for cases and controls were obtained from women who participated in Australian cancer research programs, the AOCS, the ACS, and the OPAL. Genotyping data were generated by the TRACEBACK study and the OCAC.

Ninety-eight percent of HGNMOC cases (1,544/1,578) included were known to have had germline DNA testing (in either a research or clinical setting), for at least BRCA1 and BRCA2. The extent of germline genetic testing varied, with some subjects having testing for only BRCA1/2 and others with a broader gene panel. Of tested women, 14% (212/1,544) harbored a PV, consistent with other studies (36, 53, 54). Similarly, in those with HGSOC specifically, PVs occurred in 16% (170/1,082) of tested women.

The 5- and 10-year OS rates in the cohort of 1,097 cases with HGSOC were 43% and 27%, respectively, with a median OS of 1,506 days [95% confidence interval (CI), 1,386–1,636 days]. Consistent with the literature, residual disease and stage were strongly associated with outcome, with improved survival seen in women diagnosed with an earlier stage or with no residual disease after primary surgery (Supplementary Fig. S1). In line with previous studies prior to the routine use of PARP inhibitors in the first-line setting (32), harboring a PV in an ovarian cancer susceptibility gene was associated with improved survival in the short term, but this benefit diminished over time (Supplementary Fig. S1).

PRS and distribution

The raw and standardized PRSs were calculated for cases and controls, and two datasets were generated: one for all HGNMOC cases and one for HGSOC specifically. The PRS was normally distributed and shifted to the right (higher values) in cases compared with controls (Supplementary Table S5), with no difference based on PV status (Fig. 2; Supplementary Fig. S2): all women, mean (SD) of PRS = 0.21 (1.06); excluding the PV carriers, mean (SD) = 0.20 (1.06); PV carriers only, mean (SD) = 0.24 (1.02).

Figure 2.

Distribution of the PRS in HGSOC cases and controls. A, All women; B, all women excluding PV carriers.

Assessment of overfitting and potential bias

Because a subset of participants overlapped with prior GWAS discovery datasets, sensitivity analyses were performed to assess potential overfitting and cohort-related bias. PRS distributions were compared between cases unique to this study (307 HGNMOC; 207 HGSOC) and cases previously included in GWAS discovery cohorts. No significant differences in PRS distributions were observed between unique and non-unique cases [mean (SD) = 0.29 (1.12) vs. 0.28 (1.07); P > 0.90], and the associated risk estimates were comparable (OR per SD 1.30 vs. 1.29), suggesting that the inclusion of overlapping cases did not materially inflate PRS performance (Supplementary Table S6).

The use of HGSOC-specific versus all-ovarian cancer ORs produced similar risk estimates (OR per SD, 1.40 vs. 1.29; P = 0.10), indicating no significant bias introduced by the choice of weighting.

Additionally, the potential bias related to overlap between controls and association studies was also explored by comparing observed control PRS distributions with theoretically derived population estimates based on external allele frequencies. The theoretical control mean PRS was lower than the observed control mean (2.38 vs. 2.41), resulting in a higher estimated risk using population-based expectations compared with the observed controls (OR, 1.49 vs. 1.40; P = 0.298). These findings suggest that PRS-associated risk in this cohort is unlikely to be inflated by control overlap and may instead represent a conservative estimate. Finally, no systematic differences in PRS distributions were observed between genotyping platforms, supporting consistency between TRACEBACK and OCAC-derived genotyping data (Supplementary Fig. S3).

Association of the PRS with HGSOC risk

PRS was significantly associated with HGSOC risk in this Australian population (Fig. 3). Each SD increase in PRS was associated with a 40% increase in HGSOC risk (OR, 1.40; 95% CI, 1.28–1.53; P < 0.001). Effect sizes were unchanged when restricted to cases unique to this study and were not modified by PV status (Fig. 3). Similar associations were observed in the broader HGNMOC cohort (Supplementary Fig. S4); however, exclusion of HGSOC cases attenuated the association (OR, 1.10) and resulted in a lower effect estimate (Supplementary Table S7).

Figure 3.

Association between PRS and ovarian cancer risk across analytic models. A–D, Logistic regression models demonstrating the association between PRS (continuous variable) and ovarian cancer risk, expressed as ORs with 95% CIs. Analyses include A, all cases; B, cases excluding PV carriers; C, cases unique to this study; and D, unique cases excluding PV carriers. E, Summary table of ORs, 95% CIs, and P values for each analytic group.

When analyzed categorically, women in the highest PRS quartile had increased risk compared with the middle group (OR, 1.68; 95% CI, 1.35–2.09; P < 0.001; Supplementary Fig. S5). PRS effects were again comparable in PV carriers and noncarriers, indicating an independent modifying effect of polygenic risk. PRS performance was not influenced by family history (Supplementary Fig. S6; Supplementary Table S8).

The effect of PRS on the cumulative risk of HGSOC

We assessed the clinical impact of PRS by estimating its effect on the cumulative lifetime ovarian cancer risk. Using HRs derived from the PRS and applying these across the PRS distribution to Australian population risk estimates (AIHW Cancer Data in Australia, https://www.aihw.gov.au/reports/cancer/cancer-data-in-australia), we modeled absolute lifetime risk (Fig. 4).

Figure 4.

Age-related cumulative risk of developing HGSOC (with PV carriers excluded) by PRS percentiles. Estimated cumulative lifetime risk of HGSOC across different PRS percentile groups in individuals without a known PV in established ovarian cancer susceptibility genes. Risk estimates are presented as cumulative incidence by age and demonstrate the stratification of HGSOC risk according to inherited polygenic susceptibility.

Women in the highest 1% of the PRS distribution had an estimated lifetime risk approaching 3% in the HGSOC cohort (Fig. 4) and 4% in the HGNMOC cohort (Supplementary Fig. S7), substantially higher than the average cumulative lifetime risk in Australia (1.3% AIHW Cancer Data in Australia).

Even at this extreme, the PRS-associated risk alone remained below the 5% lifetime risk threshold commonly used to guide eligibility for RRSO. However, these estimates consider PRS in isolation and do not incorporate PV status or other established risk factors. Similar patterns were observed when PV carriers were included in the analysis (Supplementary Fig. S7). Enrichment for high polygenic risk (top 1% of PRS) was similar in PV carriers (7.1%; 6/85) and noncarriers (7.5%; 36/482), further supporting the largely independent contribution of polygenic and monogenic risk.

PRS and time-to-event outcomes

We investigated the potential influence of PRS on prognosis for HGSOC, excluding PV carriers. The PRS was not associated with OS (HR = 1.03; 95% CI, 0.95–1.11; P = 0.50; Fig. 5; Supplementary Table S9). In the larger cohort of HGNMOC cases, multivariable analysis yielded consistent results, showing no association between PRS and OS (Supplementary Table S10). In light of previous literature demonstrating time-dependent effects of genomic factors on ovarian cancer outcomes (32), we evaluated whether PRS influences survival at both 5 and 10 years (Supplementary Table S11; Supplementary Fig. S8).

Figure 5.

Association between PRS and OS in HGSOC. PRS and impact on OS in HGSOC as a continuous variable in a multivariable analysis along with other prognostic factors of age, stage, and residual disease expressed as log relative hazard. Note: Residual disease variable was modeled as a stratification variable to alleviate proportional hazards assumption violation. *, A restricted cubic spline (rcs) term with four knots was used to model age at diagnosis variable.

The 5-year OS in HGNMOC cases suggested that a higher PRS may be associated with poorer survival (Supplementary Table S11). Analogous to OS, although there was a trend toward poorer survival, the remaining analyses showed no significant association with 5- or 10-year survival in the HGSOC cohort or with 10-year survival in the larger HGNMOC cohort (Supplementary Table S11). Similar findings were seen for PFS (HGSOC and HGNMOC) with no association between PRS and survival outcomes (Supplementary Fig. S9; Supplementary Tables S12 and S13).

We also assessed the potential for PRS to predict chemotherapy sensitivity in HGSOC but found no evidence of a significant association between PRS and PFI (HR= 1.04; 95% CI, 0.96–1.12; P = 0.355; Fig. 6; Supplementary Table S14). In the larger cohort of patients with HGNMOC, there was a nonsignificant trend toward PRS influencing (HR= 1.06; 95% CI, 0.99 to 1.14; P = 0.076) and potentially impeding chemotherapy sensitivity (Supplementary Fig. S10; Supplementary Table S14).

Figure 6.

PFI and PRS in HGSOC. The impact of PRS on the PFI in HGSOC using PRS as a continuous variable in a multivariable analysis along with other prognostics factor of age, stage, and residual disease expressed as a log relative hazard. Note: Residual disease variable was modeled as a stratification variable to alleviate proportional hazards assumption violation.

To examine this further and to account for the potential effect of variation in surgical outcomes, we assessed the association between the PRS and the time from diagnosis until the initiation of second- and third-line chemotherapy. The median time until second-line chemotherapy for the entire cohort was 664 days (95% CI, 590–700); 854 days (95% CI, 731–1148) in those with a PV, and 605 days (95% CI, 561–679) in those that were wild-type (WT; no PV detected). For third-line chemotherapy, it was 1,356 days (95% CI, 1,201–1,463); 1,497 days (95% CI, 1,270–2,704) in those with a PV; and 1,340 days (95% CI, 1,171–1,448) in those that were WT. Consistent with previous results, PRS was not significantly associated with time until initiation of second- or third-line chemotherapy in HGSOC using multivariable analysis (Supplementary Fig. S11; Supplementary Table S15). Similarly, time until initiation of second-line and third-line therapy was not significantly affected by PRS in the larger cohort of HGNMOC (Supplementary Fig. S12; Supplementary Table S16).

Discussion

By using a large cohort of HGSOC cases, including a set of unique cases not previously included in GWAS risk allele identification studies, we demonstrate that a higher PRS is associated with an increased risk of HGSOC in an Australian population. Women in the highest 1% of the PRS distribution had an estimated lifetime risk of approximately 3% for HGSOC and 4% for HGNMOC. The performance of the PRS in our cohort was consistent with those of previous ovarian cancer studies (22, 23), despite a comparatively lower baseline risk of ovarian cancer in this population. The association remained stable when analyses included all HGNMOC cases or were restricted to HGSOC, with no statistically significant difference between these models. In contrast, no significant association was observed in non-HGSOC cases, and the effect size was significantly attenuated, suggesting that the observed PRS signal is largely driven by HGSOC.

These findings are consistent with the biological heterogeneity of ovarian cancer and the histotype-specific nature of genetic susceptibility loci identified through genome-wide association studies. However, the lack of association observed in non-HGSOC cases should be interpreted with caution, as the smaller sample size in this subgroup may have limited statistical power to detect modest effects.

Given the absence of effective screening and the high mortality rate of HGSOC, a PRS represents a promising tool for risk stratification, with the potential to improve further as additional variants are incorporated. Moreover, with this spread of risk, if used as part of a multifactorial model including other risk factors such as family history, some women are likely to reach the 5% threshold recommended for access to risk reduction surgery that they would otherwise be ineligible for (55). Combining PRS with the germline single-gene testing also provides an estimate of how PRS modifies ovarian cancer risk associated with a PV, and our results suggest that the risks are independent.

In women with a BRCA1/2 PV, recent evidence has suggested that ovarian cancer screening led to a stage shift (a shift to diagnoses occurring in earlier stages) and was shown to be cost-effective by improving quality-adjusted life years (56). It is possible that, in the future, using PRS and ovarian cancer risk assessment tools to identify a group of at-risk women could facilitate targeted screening or early detection methods.

Despite major advances in the molecular characterization of HGSOC, including the role of HRD, which have enabled the development of personalized therapeutic strategies, the contribution of germline common variation to tumor behavior remains poorly defined. We hypothesized that SNPs captured within the PRS may influence HGSOC biology, potentially through epigenetic mechanisms, and thereby modulate chemotherapy response and survival, positioning the PRS as a candidate prognostic and predictive biomarker. However, we found no consistent evidence to support this. Although there was a suggestion of reduced platinum sensitivity and poorer survival among individuals with a higher PRS, these associations were not robust across analyses. This likely reflects the primary role of PRS in capturing disease susceptibility rather than tumor biology, compounded by limited power to detect modest prognostic effects. Future GWAS incorporating detailed treatment response and survival data will be critical to develop PRS models with meaningful predictive utility in the therapeutic setting.

In this study, several analyses were undertaken to assess potential sources of bias. We found no evidence that inclusion of cases overlapping with discovery datasets materially influenced PRS distributions or effect estimates, and comparisons with population allele frequencies suggested that any bias in the control group would likely underestimate, rather than inflate, the observed associations. Nonetheless, a residual element of overfitting due to partial overlap with discovery cohorts cannot be fully excluded.

Overall, our findings support the utility of a PRS as a tool for risk stratification in HGSOC, with potential application in identifying individuals who may benefit from targeted prevention strategies or future risk-adapted screening approaches. Future studies should focus on integrating PRS with clinical and epidemiologic risk factors and validating performance in independent and more diverse populations.

This study has several limitations. These include the limited generalizability of this PRS to non-HGSOC histotypes and to populations of non-European ancestry, reflecting the predominantly European ancestry of the GWAS datasets from which it was derived. In addition, our cohort was largely composed of individuals recorded as “White,” which was used as a proxy for European ancestry; however, race is an imperfect surrogate for genetic ancestry and does not capture underlying heterogeneity. In particular, White Australian populations may differ genetically from White European populations because of geographic variation and admixture, which may influence risk variant distribution. Additional limitations include partial overlap between this cohort and GWAS discovery datasets and the lack of clinical covariates in controls precluded adjustment for reproductive risk factors. Although germline testing status was available for most cases, the extent of testing varied and PV status was not available for controls; however, as unaffected population-based controls, the expected rate of PV in BRCA1/2 would be < 1 in 200 (57). In addition, family history data were self-reported and may be subject to misclassification.

Although supported by a biological rationale, the current PRS does not seem to have clinical utility as a biomarker for guiding therapeutic decision-making in HGSOC. However, we confirm PRS as an important contributor to HGSOC risk, with the potential to improve the accuracy of risk prediction models. Prospective validation and ongoing clinical trials (58, 59) will be critical to defining the role of PRS in clinical practice. Future models are likely to be strengthened through the integration of reproductive and other epidemiologic risk factors.

Supplementary Material

Supplementary Figure 1

Supplementary Figure 1. Kaplan Meier overall survival analyses for patients with high-grade serous ovarian cancer stratified according to disease stage, residual disease status, and germline pathogenic variant status, demonstrating differences in survival outcomes across prognostic subgroups.

Supplementary Figure 2

Supplementary Figure 2. Distribution of standardized polygenic risk score (PRS) values among individuals with and without high-grade non-mucinous ovarian cancer, illustrating differences PRS between cases and controls.

Supplementary Figure 3

Supplementary Figure 3. Comparison of standardized PRS distributions generated using different testing methodologies in the OCAC and TRACEBACK cohorts, demonstrating consistency of PRS estimates analytical approaches.

Supplementary Figure 4

Supplementary Figure 4. Association between PRS and risk of high-grade non-mucinous ovarian cancer across multiple analytical models, including analyses of all cases, exclusion of pathogenic variant carriers, unique study cases, and comparisons stratified by pathogenic variant status.

Supplementary Figure 5

Supplementary Figure 5. Univariable analysis evaluating the association between categorical PRS groups and high grade serous ovarian cancer risk, showing odds ratios and confidence intervals for low-, intermediate-, and high-risk PRS categories.

Supplementary Figure 6

Supplementary Figure 6. Performance of the polygenic risk score stratified by family history status, comparing PRS distributions and ovarian cancer risk between individuals with and without a reported family history.

Supplementary Figure 7

Supplementary Figure 7. Age-related cumulative ovarian cancer risk stratified by polygenic risk score categories in high-grade non-mucinous ovarian cancer excluding pathogenic variant carriers and in high-grade serous ovarian cancer including pathogenic variant carriers, demonstrating differences in cumulative risk across PRS groups.

Supplementary Figure 8

Supplementary Figure 8. Multivariable analyses evaluating the association between PRS and overall survival in high-grade serous ovarian cancer at 5-year and 10-year follow-up, adjusted for age, disease stage, and residual disease status.

Supplementary Figure 9

Supplementary Figure 9. Multivariable analysis assessing the relationship between PRS and 5-year progression-free survival in high-grade serous ovarian cancer while accounting for age, stage, and residual disease as prognostic factors.

Supplementary Figure 10

Supplementary Figure 10. Multivariable analysis examining the association between PRS and platinum-free interval in high-grade non-mucinous ovarian cancer, adjusted for age, stage, and residual disease status.

Supplementary Figure 11

Supplementary Figure 11. Association between PRS and time to subsequent chemotherapy in high-grade serous ovarian cancer, including analyses of time to second-line and third-line chemotherapy adjusted for age, stage, and residual disease.

Supplementary Figure 12

Supplementary Figure 12. Association between PRS and time to subsequent chemotherapy in high-grade non-mucinous ovarian cancer, including analyses of time to second-line and third-line chemotherapy while accounting for clinical prognostic factors.

Supplementary Tables 1-16

Supplementary Table 1: Clinical data collected for ovarian cancer patients. Supplementary Table 2: Ovarian cancer susceptibility SNPs included in the PRS. Supplementary Table 3: SNP performance and selection for the final PRS. Supplementary Table 4: Clinical characteristics of the study cohort. Supplementary Table 5: HGSOC case–control summary statistics. Supplementary Table 6: Comparison of PRS in unique and non-unique cases. Supplementary Table 7: Association between PRS and ovarian cancer risk in HGNMOC and HGSOC. Supplementary Table 8: PRS distribution according to family history of ovarian cancer. Supplementary Table 9: Association of PRS with overall survival in HGSOC. Supplementary Table 10: Association of PRS with overall survival in HGNMOC. Supplementary Table 11: Association of PRS with 5-year and 10-year overall survival. Supplementary Table 12: Association of PRS with 5-year and 10-year progression-free survival. Supplementary Table 13: Association of PRS with 5-year progression-free survival in HGSOC. Supplementary Table 14: Association of PRS with 10-year platinum-free interval. Supplementary Table 15: Association of PRS with time to second- and third-line chemotherapy in HGSOC. Supplementary Table 16: Association of PRS with time to second- and third-line chemotherapy in HGNMOC.

Data Availability

Genotyping in this study was either (a) generated by the TRACEBACK study or (b) generated previously by OCAC. Raw data and derived PRS data supporting the findings of this study can be requested from the originating studies with an ethically approved project (aocs@petermac.org; traceback@petermac.org).

Authors’ Disclosures

R. Delahunty reports grants from Australian government and Tour de Cure and other support from National Health and Medical Research Council (NHMRC) and Victorian Comprehensive Cancer Centre Alliance during the conduct of the study, as well as other support from MSD outside the submitted work. D.W. Garsed reports grants from NHMRC (Australia), US Army Medical Research and Materiel Command Ovarian Cancer Research Program, and Victorian Cancer Agency during the conduct of the study. S. Fereday reports grants from AstraZeneca Pty Ltd. during the conduct of the study. N. Traficante reports grants from AstraZeneca Pty Ltd. during the conduct of the study. E.L. Christie reports grants from AstraZeneca Pty Ltd. outside the submitted work. P.M. Webb reports grants from NHMRC (Australia), US Army Medical Research and Materiel Command, and Cancer Council Victoria, Cancer Council Queensland, Cancer Council New South Wales, Cancer Council South Australia, Cancer Council Tasmania, and Cancer Foundation of Western Australia during the conduct of the study. K. Alsop reports grants from AstraZeneca Pty Ltd. outside the submitted work. P.A. James reports grants from NHMRC (Australia) during the conduct of the study. No disclosures were reported by the other authors.

Authors’ Contributions

R. Delahunty: Conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, writing–original draft, writing–review and editing. S.S.M. Silva: Data curation, formal analysis, investigation, methodology, writing–review and editing. A. Pandey: Software, formal analysis, investigation, methodology, writing–review and editing. D.W. Garsed: Data curation, methodology, writing–review and editing. J. Beesley: Resources, data curation, investigation, methodology, writing–review and editing. S. Fereday: Data curation, investigation, project administration, writing–review and editing. N. Traficante: Data curation, investigation, methodology, project administration, writing–review and editing. S. Craig: Data curation, investigation, writing–review and editing. L.N. Nguyen: Data curation, investigation, project administration, writing–review and editing. J. Hendley: Data curation, investigation, writing–review and editing. E.L. Christie: Data curation, supervision, methodology, writing–review and editing. P.M. Webb: Resources, data curation, investigation, methodology, writing–review and editing. G. Chenevix-Trench: Resources, data curation, investigation, methodology, writing–review and editing. D.D.L. Bowtell: Conceptualization, resources, supervision, funding acquisition, writing–review and editing. K. Alsop: Data curation, investigation, methodology, project administration, writing–review and editing. M. Jayawardana: Data curation, formal analysis, investigation, methodology, writing–review and editing. P.A. James: Conceptualization, supervision, funding acquisition, methodology, writing–review and editing.