Age-adjusted association of homologous recombination genes with ovarian cancer using clinical exomes as controls

Next-generation sequencing (NGS) has enabled clinical laboratories to analyze simultaneously a growing number of genes. Clinical multi-gene hereditary ovarian cancer (OC) panels include genes that function in the same homologous recombination (HR) DNA repair pathway as BRCA1 and BRCA2, such as ATM, BARD1, BRIP1, CHEK2, NBN, PALB2, RAD51C, and RAD51D. These genes have been linked to hereditary ovarian cancer, but the extent to which some of these genes contribute to hereditary OC varies in the literature [1‐7]. The identification of pathogenic or likely pathogenic variants (PVs) in these genes with ambiguous associations leads to difficult patient management decisions for clinicians and patients alike.

Identifying an appropriate control group for genetic association studies is vital for accurate estimation of cancer risks. Before high-throughput NGS, researchers relied on collaboration by pooling multi-center genotype data to achieve a large enough sample size to detect significant genetic associations [8, 9]. Recent association studies have drawn control data from the Exome Aggregation Consortium (ExAC), a public database which contains genotype information on over 60,000 individuals who have participated in genetic studies on conditions such as inflammatory bowel disease, heart disease, and schizophrenia [10‐12]. The FLOSSIES database has also been used to perform genetic association studies specifically in women’s cancer [13]. The advantage of such cohorts are the large, readily available number of genotyped individuals with diverse ancestries and varied clinical histories, but drawbacks include the lack of specific phenotypic information including age and cancer history of the participants.

Clinical laboratories are addressing this limitation by leveraging data from exome sequencing (ES) tests performed internally [14]. In our laboratory, ES can be performed on trios (child and both parents) to help identify variants contributing to rare Mendelian diseases. Healthy parents from these trios serve as a reliable control cohort for hereditary cancer studies considering the breadth of genotyping coverage and the adequate phenotypic information provided. We performed a case-control study with an internal control population and report age-adjusted ORs to clarify equivocal genetic associations in HR genes with OC.

Our approach is supported by the agreement of OR_Crude and OR_Adj with previously reported ORs for BRCA1, BRCA2, RAD51C, and RAD51D, genes with clearly established OC associations. The other HR pathway genes have been reported with conflicting evidence for association with OC [10, 14, 23‐25]. Association of PVs in each these genes with OC are in agreement with at least one of these studies by either OR_Crude or OR_Adj. While this validates that our methodology is comparable, it also highlights the disparity of concordance among these studies.

To check for potential ascertainment bias, PV prevalence in cases and controls was compared to those previously reported by Lilyquist et al. [26] Given the similar ascertainment of cases and robustness of the estimates from the ExAC non-Finnish European control population, this data serves as a germane baseline for comparison. The prevalence of PV was not significantly different in any gene (Additional file 6: Table S5). Small differences in PV counts due to differences in variant classification between studies can have a significant impact on effect size, especially for genes where PVs are rare. For example, Weber-Lassalle et al. reported 9 of 7,325 (0.12%) European PV carriers from the FLOSSIES database; however, our variant classification system would have reported 14 of 7,325 (0.19%), on the basis of including 5 European carriers of BRIP1 c.139C > A (ClinVar: SCV000210833.12) (Fig. 1) [13].

Genes in which PVs are rare in one cohort (four or less PVs), BARD1, BRCA1, NBN, RAD51C and RAD51D, result in wide confidence intervals for the ORs. For example, RAD51C appears to confer greater risk than BRCA2, a gene whose association with OC has been well-established. However, with the exception of CHEK2, confidence intervals for ORs in all HR genes overlapped with those previously published, indicating the true effect size falls somewhere in the overlapping range [10]. Results from the power analysis can be interpreted that genes with moderate effect sizes are adequately powered to detect significant association but the result also supports the argument in favor of performing studies on genotyped cohorts large enough to sufficiently detect genetic associations in genes where PVs rare and the effect sizes are small.

No significant association with OC were consistently observed with PVs in BARD1, NBN, and PALB2. The reported low frequency of PVs in BARD1, similar to other studies, will require still larger sample sizes to detect a significant effect [10, 23, 24]. Association of NBN with OC is the low to moderate risk HR pathway gene with the most concordant results. It is widely regarded as not significantly associated with OC, with the exception of Lilyquist, et al. who reported that its association was “marginally significant” [23]. Additionally, Lilyquist et al. initially reported a significant association of PALB2 with OC, but the association was lost upon removal of women with a personal or family history of breast cancer [23]. We also report no significant association of PALB2 with OC.

Multiple studies have reported ATM as a moderate risk OC gene [14, 23]. The observed significance by OR_Crude is concordant with these results. In controls, ATM PV carriers had the highest median age (47 years, time of testing) of any HR pathway genes. Given that PVs in ATM were only moderately associated with OC by OR_Crude, the median age of PV carriers was high enough compared to the median age of non-carriers to decrease the OR_Adj below the significance threshold. Notably, an insignificant z-test result indicates that the true effect size is likely to overlap the two reported CIs. The presence of undiagnosed cancer patients in the control cohort could falsely lower the calculated ORs, a similar effect was previously described in a study which used the ExAC controls with cancer samples included [10].

A well-known breast cancer risk gene, CHEK2, has been consistently reported to have no significant association with OC [10, 14, 23]. Similar to reported associations from ATM, the observed significance by OR_Crude is concordant with previously reported associations. In contrast, however, CHEK2 PV carriers in cases had the lowest median age (52 years, time of OC diagnosis) of any of the HR pathway genes, which increased the OR_Adj enough to reach the level of significant association with OC. Again, similar to ATM, the z-test comparing the two effect sizes was insignificant, indicating that the true effect size is likely in the overlapping range of the CIs (Table 1). When limiting cases to women with serous OC pathology, association of OC with PVs in CHEK2 disappears and the median age of OC diagnosis in PV carriers increases from 52 years to 63.5 years (Additional file 5: Table S4). None of the 12 CHEK2 PV carriers with serous OC subtype were also diagnosed with breast cancer (not shown), suggesting that the significant OR_Adj result from the larger case cohort was a direct effect from the bias introduced by the younger CHEK2 PV carriers who were also diagnosed with breast cancer.

Multiple publications have concluded that BRIP1 is significantly associated with OC [10, 13, 23, 24]. While OR_Crude in BRIP1 was significantly associated with OC, the OR_Adj showed no OC association. The significantly decreased effect size in BRIP1 can be attributed to PV carriers displaying the oldest median age at time of OC diagnosis and the second oldest median age at time of testing among all of the genes (67y case PV carriers, 46y control PV carriers; Table). Advanced age of PV carriers in BRIP1 is not a novel observation [10, 13, 23, 24]. BRIP1 was the only gene in which OR_Adj was significantly different than OR_Crude, which, given PV carriers are older compared to non-carriers, suggests that BRIP1 PV carriers are more likely to be diagnosed with late-onset OC. Recent updates to National Comprehensive Cancer Network® (NCCN) guidelines for BRIP1 PV carriers include consideration of risk-reducing salpingo-oophorectomy at 45–50 years of age [26]. The consistently reported older age at time of diagnosis and the observed lack of association by OR_Adj suggests caution before surgical intervention and the need for further studies of larger cohorts of older controls, as previously recommended [27].

A recent study reported no significant association of BRIP1 with OC, but our results reveal a clinically relevant factor that offers insight which may have contributed to the observed lack of association. Age (mean [SD]) at time of OC diagnosis in cases (55.7y [14.1]) and age at time of testing in controls (39.7y [14.7]) were both younger than our cohorts, which could lead to under-counting BRIP1 PV carriers. Finally, it has been suggested that BRIP1’s OC association may be restricted to high-grade serous epithelial ovarian cancer, but when cases were restricted to women diagnosed with high-grade serous OC histologic subtype, the OC association was consistent (Additional file 5: Table S4) [7].

By presenting OR_Adj in addition to OR_Crude this study allows for comparison of the two effect sizes. This comparison provides insight into the age difference between PV carriers and non-carriers and enables inference of early/late onset OC. Genes with PV carriers who are older than non-carriers demonstrate decreased OR_Adj compared to OR_Crude and conversely, genes with PV carriers who are younger than non-carriers have increased OR_Adj compared to OR_Crude. Comparing the two effect sizes for each gene using a z-test revealed that controlling for age did not significantly change the OR, except in BRIP1. A significant z-test result suggests that BRIP1 PV carriers are more likely to be diagnosed with late-onset OC (Table 1).

As a referral laboratory, clinical information was limited to that provided by ordering providers with submitted samples. With ES, the referrals are not routinely submitted for cancer-related testing and, therefore, cancer history may not have been included in the clinical histories by the ordering physician. In addition, because the mean age of controls was lower than the mean age of OC diagnosis, it cannot be ruled out that some of these women will ultimately be diagnosed with OC. The OC referral cases could be biased toward a higher risk than the general ovarian cancer population. Variant detection sensitivity filters for controls were conservatively chosen which may lead to under-reporting of PVs and thus overestimation of ORs. Another source of possible under-reporting of PVs in controls is due to the incomplete gene coverage in controls, although the minor difference would have negligible effect (Additional file 2: Additional Sequencing Methods and Table S2). Copy number variants would have likely contributed a small number of PVs to cases and controls. But since this type of variant was not evaluated it acts as a source of bias toward under-reporting PVs. As not all PVs in controls were orthogonally confirmed as they were in cases, it is possible that a small number of control PVs were sequencing artifacts. Targeted BRCA1 and BRCA2 testing tends to have a lower positive yield compared to panel testing with other HR genes, as previously described [28]. The overall pathogenic variant rate in OC (9.5%) is lower than published rates from other studies [10, 23]. This can be attributed 873 women in the case cohort who underwent testing for BRCA1 and/or BRCA2 only.

We present age-adjusted genetic associations for PVs in HR genes with OC, leveraging an internal control cohort of women who self-report as White/Caucasian. Our study design and analysis provide more informed estimates of association compared to recently published OC associations by reporting both OR_Crude and OR_Adj. These results are most relevant for BRIP1 PV carriers, as our findings for this gene are disparate from recent literature and conflict with current management guidelines.