Background

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 [1,2,3,4,5,6,7]. 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 [10,11,12]. 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.

Methods

Cohort assembly

Genotypic and phenotypic data were collected from women who underwent germline genetic testing at GeneDx (Gaithersburg, MD) between 2013 and 2018. The study was conducted in accordance with all guidelines set forth by the Western Institutional Review Board, Puyallup, WA (WIRB 20162523). Informed consent for genetic testing was obtained from all individuals undergoing testing, and WIRB waived authorization for use of de-identified aggregate data for both cases and controls. Individuals or institutions who opted out of this type of data use were excluded. Cases were women with OC referred for BRCA1/2 alone or multi-gene hereditary cancer panel testing who did not also have a targeted test for a known familial variant. Controls were mothers from a subset of complete trios referred for ES due to a neurodevelopmental delay in the proband. These mothers all self-reported that they did not have a disorder with genetic etiology. Both case and control cohorts were limited to those over 18 years of age and those who self-reported as White/Caucasian. Women were excluded if they were missing age at diagnosis or age at testing or if they had more than one PV, including homozygous variants, in any of the cancer predisposition genes available on ordered panels (Additional file 1: Table S1). Of all individuals referred for multi-gene hereditary cancer panel testing, 9,688 women were diagnosed with OC, 6,182 of whom met inclusion criteria. From the control cohort, there were 8,643 mothers of probands, of whom 4,690 satisfied the filtering criteria.

Sequencing and variant calling

Cases were sequenced and genotyped with targeted NGS panels and a clinical bioinformatics pipeline as previously described [15]. Controls were sequenced by clinical ES as previously described using either Agilent SureSelect Human All Exon v4 or Agilent Clinical Research Exome capture protocols (Agilent, Santa Clara, CA) [16]. Separate from clinical ES testing, control samples were jointly genotyped across the entire cohort following Genome Analysis Toolkit Best Practices using HaplotypeCaller (version 3.7.0) in GVCF mode followed by GenotypeGVCFs and variant quality score recalibration [17,18,19]. Single nucleotide variants with genotype quality less than 50 and insertions and deletions (indels) with genotype quality less than 99 were considered low confidence, and therefore not included in the analysis. Additionally, for all novel prospective PVs in controls, de-identified NGS alignment data were manually inspected to remove suspicious variants. Sequencing methods for cases and controls were evaluated for potential bias (Additional file 2: Additional Sequencing Methods and Table S2). Copy number variants were not evaluated as part of this study.

Variant classification

Analysis of genetic variants were limited to those located in the HR genes, ATM, BARD1, BRCA1, BRCA2, BRIP1, CHEK2, NBN, PALB2, RAD51C, and RAD51D. For cases, the pathogenic classification of each variant was reviewed according to internal protocol, which follows ACMG/AMP guidelines [20]. Exon-level deletions and duplications were not included in the PV count since detection of these copy-number variants from ES is more limited and variable than from panel testing. Pathogenic status of variants in controls were systematically classified using a rule-based algorithm (Fig. 1). Variants that were previously classified in cases were assigned the same classification. Novel variants in the controls were assigned ClinVar classifications or were manually classified when ClinVar classifications were not available or conflicting. ClinVar classifications were obtained from the variant call format file released September 5, 2017 [21]. The common founder pathogenic missense variants, CHEK2*I157T and CHEK2*S428F were excluded from both cases and controls. All PVs are listed in Additional file 3: Table S3.

Fig. 1
figure 1

Rule-Based Algorithm Used to Classify Variants. #No classification discrepancies were identified for variants falling into this category (not previously seen at GeneDx but in ClinVar). ^Variants requiring manual review were not classified based on strict ACMG criteria as is done for variants that were clinically reported. Variants were classified as either a PV, which includes variants that would meet criteria for a pathogenic or likely pathogenic classification, or not a PV. Variants determined to be not a PV were not worked up further to determine if they would be classified as variant of uncertain significance, likely benign, or benign. *NGS data manually assessed to ensure variant was real. Default as real if it was determined that the variant in question could not be confidently called real or not real based on NGS data. NMD = nonsense mediated decay

Statistical analysis

A two-sample, independent t-test was performed to compare the mean age between cases and controls. Genetic associations were estimated by ORs using Firth bias-corrected logistic regression [22]. Crude ORs (ORCrude) were generated by simple logistic regression using status of PV in a gene as the independent variable. Multivariable logistic regression was performed with PV status and patient age (diagnosis in cases, at time of testing in controls) as independent variables to generate ORAdj. Reported 95% confidence intervals (CIs) were calculated using penalized profile likelihood. A z-test was performed to compare ORCrude and ORAdj for each gene (Additional file 4). The aggregate PV prevalence per-gene was compared to reported prevalence from Lilyquist et al. using Fisher’s exact test [23]. To correct for multiple independent tests, a Benjamini-Hochberg false discovery rate (FDR) correction was applied and a corrected p-value significance threshold of 0.05 was used when considering genetic association results and results from Fisher’s exact test. The t-test and z-tests were two-sided and P < 0.05 were considered significant. A sensitivity analysis was performed by reporting OC associations in a subset of women who were diagnosed with serous OC subtype (Additional file 5: Table S4). Finally, a power analysis was performed using a test of proportions to determine the statistical power of the study. Conditions of the test used the total cohort sizes, assumed a significance level of 0.05, and a subjective, but realistic, assumption that rate of PVs in each cohort were 0.01, and 0.005, respectively. Statistical analyses were performed using the R programming language.

Results

Cohort summary

After filtering into final cohorts, the mean age (standard deviation) at OC diagnosis was significantly higher in cases than age at testing in controls (58.8 (13.4) versus 41.9 (9.4), P < .05, t-test). The median age at diagnosis of PV carriers in OC cases ranged from 53 years in BRCA1 to 67 years in BRIP1, while the median age of testing of PV carriers in controls ranged from 32.5 years in RAD51D to 46 years in BRIP1 (Table 1). There were 780 (12.6%) OC cases who also had a breast cancer diagnosis. Across all genes, the observed aggregate PV prevalence in either the case or control cohort were not significantly different, respectively, from those reported by Lilyquist et al. (P = 0.05, Fisher’s exact test). There were 588 women of the 6,182 cases who harbored PVs (9.5%) while 98 of the 4,690 controls (2.1%) harbored PVs (Table 1).

Table 1 Associations of ovarian cancer with pathogenic variants in homologous recombination genes among self-reported white women

Genetic associations

Well-established OC susceptibility genes showed significant associations when measured by both ORCrude and ORAdj: BRCA1 (ORCrude = 38.46, CI = 17.03 to 115.32, P = 1.03 × 10− 47; ORAdj = 47.80, CI = 20.76 to 145.03, P = 1.87 × 10− 46), BRCA2 (ORCrude = 8.19, CI = 5.22 to 13.67, P = 1.17 × 10− 28; ORAdj = 5.10, CI = 3.12 to 8.83, P = 2.89 × 10− 12), RAD51C (ORCrude = 15.28, CI =3.97 to 137.01, P = 6.09x− 06; ORAdj = 12.09, CI = 2.78 to 114.43, P = 7.85 × 10− 04), and RAD51D (ORCrude = 7.99, CI = 2.56 to 39.77, P = 1.97 × 10− 04; ORAdj = 8.38, CI =2.17 to 47.23, P = 2.51 × 10− 03) (Table 1).

Some genes with no significant association with OC measured by ORCrude consistently showed no significant association with OC as measured by ORAdj, including BARD1 (ORCrude = 1.59, CI = 0.31 to 9.56, P = 0.57; ORAdj = 6.30, CI =0.55 to 74.25, P = 0.19), NBN (ORCrude = 0.52, CI = 0.15 to 1.54, P = 0.27; ORAdj = 0.45, CI = 0.10 to 1.82, P = 0.33), and PALB2 (ORCrude = 2.06, CI = 0.90 to 5.19, P = 0.11; ORAdj = 1.78, CI = 0.61 to 5.59, P = 0.33).

ATM (ORCrude = 2.01, CI = 1.15 to 3.64, P = 0.03; ORAdj = 1.72, CI = 0.89 to 3.44, P = 0.18) and BRIP1 (ORCrude = 2.05, CI = 1.11 to 3.94, P = 0.03; ORAdj = 0.87, CI =0.41 to 1.93, P = 0.73) were significantly associated with OC by ORCrude but were not significantly associated with OC by ORAdj. Conversely, CHEK2 (ORCrude = 1.62, CI = 1.00 to 2.69, P = 0.07; ORAdj = 2.64, CI = 1.48 to 4.79, P = 2.36 × 10− 3) was not significantly associated with OC by ORCrude but was significantly associated with OC by ORAdj.

Power analysis and effect size comparison

The power analysis revealed that the study can detect a difference in proportions from 0.01 to 0.005 with 85.9% power. The per-gene comparison of effect sizes was not significantly different for the majority of the genes evaluated. The single exception was BRIP1, for which ORAdj was significantly lower than ORCrude (z-test, P = 0.04) (Table 1).

Discussion

Our approach is supported by the agreement of ORCrude and ORAdj 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, 23,24,25]. Association of PVs in each these genes with OC are in agreement with at least one of these studies by either ORCrude or ORAdj. 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 ORCrude 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 ORCrude, the median age of PV carriers was high enough compared to the median age of non-carriers to decrease the ORAdj 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 ORCrude 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 ORAdj 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 ORAdj 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 ORCrude in BRIP1 was significantly associated with OC, the ORAdj 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 ORAdj was significantly different than ORCrude, 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 ORAdj 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 ORAdj in addition to ORCrude 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 ORAdj compared to ORCrude and conversely, genes with PV carriers who are younger than non-carriers have increased ORAdj compared to ORCrude. 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.

Conclusions

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 ORCrude and ORAdj. 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.