A BRCA1/2 Mutational Signature and Survival in Ovarian High-Grade Serous Carcinoma

Advances in sequencing technology and informatics have led to the identification of mutational signatures in human cancer (1). These signatures are defined by the type and frequency of somatic events and represent a molecular phenotype of the underlying etiologic factors involved in cancer development.

Patterns of somatic sequence alterations have been associated with specific tumor types and mutational processes. The most common mutational signature in human cancer shows an enrichment of C>T transitions at CpG sites, which reflect spontaneous deamination of methylated cytosine to thymine and has been associated with age. Analysis of breast and ovarian carcinomas has identified a distinct mutational signature associated with loss of function mutations in BRCA1 or BRCA2 (1, 2). Genomes of cancers with BRCA1/2 mutations have a relatively even distribution of types of nucleotide substitutions and lack of enrichment of C>T transitions compared to BRCA1/2 wild-type cancers.

The identification of BRCA1/2 mutations in breast and ovarian cancer has been increasingly recognized to have clinical significance beyond implications for hereditary cancer risk. BRCA1/2 mutations are associated with response to platinum-based chemotherapy and favorable outcome in ovarian cancer (3–6). Tumor cells with BRCA1/2 defect show in vitro sensitivity to PARP inhibitors (7, 8), and the use of PARP inhibitors to treat cancers with BRCA1/2 mutation is a subject of clinical trials (9, 10).

No prior studies have evaluated the prognostic significance of mutational signatures in ovarian cancer. In this article, we develop a Bayesian statistical model to identify ovarian cancers that possess a BRCA1/2 mutational signature, and we evaluate the prognostic significance of this classification in two independent cohorts of high-grade ovarian serous carcinoma.

The Cancer Genome Atlas (TCGA) open access whole exome sequencing and clinical data from 310 ovarian serous carcinomas was obtained from the TCGA data portal (https://tcga-data.nci.nih.gov/tcga/). Sixteen tumors with less than 10 somatic mutations were excluded from the Bayesian analysis. Germline and somatic BRCA1 and BRCA2 mutations and somatic copy number alterations were obtained via the TCGA ovarian cancer publication and cBioPortal for Cancer Genomics (http://www.cbioportal.org/; refs. 3, 11). All samples with BRCA1 or BRCA2 variants were classified as BRCA1/2 mutated, with the exception of samples with only BRCA2 K3326*, a known benign variant. Clinical follow up was available for 292 of 294 patients.

The Australian Ovarian Cancer Study (AOCS) open access whole genome sequencing and clinical data from 92 donors was obtained from the International Cancer Genome Consortium data repository (https://dcc.icgc.org/; ref. 12). For donors with multiple samples, only one sample with the lowest sample identification code was used for analysis.

Single nucleotide substitutions were mapped to human genome version 37 via Galaxy (https://usegalaxy.org/), and the 5′ and 3′ nucleotides adjacent to the substitution event were extracted. Consideration of six substitution classes (C>A, C>G, C>T, T>A, T>C, T>G) and the adjacent bases accounted for 96 possible substitution types. For each tumor, we computed the proportion of total substitutions from each of the 96 substitution types, and we used these proportions as the 96 input features for the Naïve Bayes analysis.

We used Naïve Bayes to build a classifier to compute the conditional posterior probability of BRCA1/2 mutational status given the observed proportions of mutational substitution types. To implement the analysis, we used the NaiveBayes function in the klaR package in R with default parameters. On the TCGA dataset, we built the model using mutational status, including germline and somatic BRCA1/2 mutations (classification label), and the mutational substitution proportions (input features), and we predicted for each tumor the presence or absence of a BRCA1/2 mutation, which we defined as a probabilistic representation of the tumor's mutational signature. The probabilistic prediction was then converted into a binary variable by thresholding at a posterior probability cut-off of 0.5, and tumors with probability greater than 0.5 were defined as having a BRCA1/2 mutational signature. The binary BRCA1/2 mutational signatures were then used in survival analyses. On the AOCS dataset, we applied the Bayesian model learned on the TCGA dataset to the mutational substitution proportions from the AOCS tumors to predict the presence or absence of a BRCA1/2 mutational signature. The binary BRCA1/2 signature predictions were then used in survival analyses.

Chi-square test was used to test for differences in frequencies of mutational substitution types. Mann–Whitney test was used to compare distributions of continuous variables. The log-rank test was used for the comparison of survival distributions. Univariate and multivariate survival analyses were performed using Cox proportional hazard models. All tests were two-sided with statistical significance set at P < 0.05. Bonferroni correction was applied to adjust for multiple comparisons, where appropriate.

A total of 15,439 somatic substitution mutations from 310 ovarian serous carcinomas were classified into 96 subtypes based on six substitution classes (C>A, C>G, C>T, T>A, T>C, T>G) and the flanking 5′ and 3′ nucleotides. A total of 69 of 310 carcinomas (22.3%) exhibited germline or somatic mutations in BRCA1/2. BRCA1/2 mutated serous carcinomas demonstrated distinct mutational signatures compared to BRCA1/2 wild-type carcinomas (Fig. 1 and Supplementary Table S1).

In ovarian high-grade serous carcinomas, the most common mutation types involved C>T transitions at CpG sites, which accounted for 16.2% of all mutations. BRCA1/2 mutated carcinomas exhibited a decreased enrichment of C>T transitions at CpG sites compared to BRCA1/2 wild-type carcinomas (11.4% vs. 17.9%, P ≤ 0.0005). In addition, BRCA1/2 mutated carcinomas showed a significant relative increase of T>A transversions at GpTpC sites (0.82% vs. 0.33%, P = 0.0002).

A Bayesian model was constructed and applied to each tumor to derive a posterior probability of BRCA1/2 mutation status based on the tumor's overall mutational substitution profile. To gain insight into the mutational patterns driving the model's predictions, we evaluated the mutational profiles of tumors with high confidence associations with posterior probability less than 0.1 (least BRCA1/2-like) or greater than 0.9 (most BRCA1/2-like). These tumors were shown to be particularly enriched for differentiating mutational subtypes. As expected, the mutational profiles of tumors predicted to have a low probability of a BRCA1/2-associated mutational signature were found to have increased prevalence of C>T transitions at CpG sites (Fig. 2). These findings supported that the Bayesian predictor classified cancers based on mutational signatures.

Although the predicted BRCA1/2 mutational signature was highly correlated with the presence of BRCA1/2 mutation, this association was imperfect, showing an area under the receiver operating characteristic curve of 0.86 ± 0.02 (Fig. 3A–C). The Bayesian prediction and BRCA1/2 mutational status were discordant in 74 of 294 tumors (25%) with sensitivity of 80.0% and specificity of 73.4% (Supplementary Table S2). This analysis suggested that a subclass of ovarian serous carcinomas with BRCA1/2 mutations had mutational signatures similar to BRCA1/2 wild-type carcinomas, and a subclass of ovarian serous carcinomas without BRCA1/2 mutations had signatures similar to BRCA1/2-mutated carcinomas.

We postulated that the presence of a BRCA1/2 mutational signature may provide additional information to BRCA1/2 mutational status alone to more precisely subtype ovarian high-grade serous carcinoma. To begin to evaluate this hypothesis, we studied the association between BRCA1/2 mutation, BRCA1/2 mutational signature, age, and mutational rate.

Patients with BRCA1/2 mutation were diagnosed at a younger age (median 55 years compared to 61 years for patients without BRCA1/2 mutation, P = 0.003). Among BRCA1/2 wild-type patients, the median age at presentation of patients with a BRCA1/2 mutational signature was similar to patients with wild-type signature (P = 0.29; Fig. 3D). Total mutational burden was positively associated with the BRCA1/2 mutational signature overall (P < 0.0001) and in analyses stratified by BRCA1/2 mutational status (both P ≤ 0.01; Fig. 3E). These findings suggested that the BRCA1/2 mutational signature was not significantly associated with age but was associated with biological features of cancer development such as mutational burden.

In a subset of 62 BRCA1/2 mutated carcinomas with copy number information, copy number loss of the gene with sequence alteration (BRCA1 or BRCA2) was compared to mutational signature. The majority of neoplasms (50 of 62, 81%) displayed copy number loss in BRCA1 or BRCA2 together with a mutation. A total of 44 of 50 carcinomas (88%) with BRCA1/2 mutational signature showed copy loss of BRCA1/2, whereas 6 of 12 carcinomas (50%) with wild-type signature showed BRCA1/2 copy loss (P = 0.01).

We then evaluated the association of BRCA1/2 mutations and mutational signatures with clinical outcome. Most patients in the TCGA cohort had clinical stage III–IV and pathologic grade 3 cancers and were treated with adjuvant chemotherapy without radiotherapy (Supplementary Table S3).

Both BRCA1/2 mutation and the BRCA1/2 mutational signature were associated with improved overall survival in univariate analyses (Fig. 4A and B, both P ≤ 0.0002). Median survival was 55.2 months for patients with a BRCA1/2 signature compared to 38.0 months for patients with a wild-type signature. The BRCA1/2 signature remained significant in analyses stratified by BRCA1/2 mutation status (Fig. 4C and D). These findings showed that the BRCA1/2 mutational signature had independent prognostic value in ovarian cancer beyond that achieved by BRCA1/2 mutational status alone.

We evaluated the prognostic value of the BRCA1/2 signature in multivariate survival analysis by a Cox proportional hazards model, which also considered BRCA1/2 mutation, age, stage, and grade. In this model, BRCA1/2 mutation status (HR = 0.58; P = 0.02) and BRCA1/2 mutational signature (HR = 0.64; P = 0.02) were statistically significant and independent predictors of improved overall survival (Table 1).

Finally, we validated the prognostic value of the BRCA1/2 mutational signature in an independent cohort from the Australian Ovarian Cancer Study. Here we considered an additional 5,757 mutations from 92 ovarian high-grade serous carcinomas. The mutational frequencies from BRCA1/2-mutated and wild-type carcinomas in TCGA were used to calculate Bayesian posterior probabilities of each cancer in AOCS.

The BRCA1/2 mutational signature was associated with improved survival in the AOCS cohort in univariate analyses (Fig. 4E; P = 0.003). Median survival was 46.3 months for patients with a BRCA1/2 signature compared to 23.6 months for patients with a wild-type signature in the AOCS cohort. The BRCA1/2 signature remained significant in multivariate models adjusted for stage and age (HR = 0.52; P = 0.007; Table 2).

In this study, we develop a Bayesian statistical model using somatic substitution mutation frequencies and identify a subset of ovarian high-grade serous carcinomas with a distinctive mutational signature associated with BRCA1/2 mutation. The BRCA1/2 mutational signature is present in some patients without demonstrable BRCA1/2 mutation and is an independent predictor of outcome along with BRCA1/2 mutational status in multivariate analysis.

Breast and ovarian cancers with BRCA1/2 mutation have been well characterized to have unique biological and clinical features, collectively termed “BRCAness” (13). Features of BRCAness include DNA repair deficiency, distinctive gene expression patterns, response to platinum-based chemotherapy, and sensitivity to PARP inhibition. The field has dedicated significant effort to identifying BRCA1/2-associated features in sporadic cancers with the hope of expanding therapeutic options for a greater number of patients. Prior work has evaluated BRCAness in sporadic ovarian cancer by gene expression profiling and by analysis of somatic genomic alterations, including genomic instability, loss of heterozygosity, and mutational burden (14–17). These studies measure a variety of biological outputs and consistently identify BRCAness in a subset of sporadic cancers with implications in drug response and prognosis. The mechanism of BRCAness in ovarian cancers without BRCA1/2 mutation is not well understood, but studies have implicated mutations in other homologous recombination genes in pathogenesis (18).

More recently, the broad sequencing of cancer genomes have revealed mutational signatures in human cancer (1). Signatures of somatic alterations reflect mutational processes during tumor development, including a BRCA1/2-associated signature in breast and ovarian cancer. In that study, non-negative matrix factorization is used to approximate the contribution of a discrete number of signatures to the overall mutational profile of a cancer (19). Compared to non-negative matrix factorization, our Bayesian model also considers the overall mutational profile and uses the presence or absence of a known driver mutagenic exposure (e.g., BRCA1/2 mutational status) to supervise model construction. This method may be readily applied to large sequencing datasets for hypothesis testing to derive and evaluate mutational signatures of a variety of known genetic and environmental exposures implicated in carcinogenesis.

We utilize open access data from two large cohorts to make observations about prognosis, and our analysis has limitations. The observed mutational signatures are associated with BRCA1/2 mutations, but the rationale for the presence of a BRCA1/2 mutational signature in the absence of mutation in some tumors is not explained in this analysis. These cases may reflect alternative mechanisms of BRCA1/2 inactivation, defects of other genes in the homologous repair pathway, or limitations of whole exome sequencing in the TCGA dataset to identify BRCA1/2 mutations. Similarly, a subset of patients with BRCA1/2 mutation may have nonpathogenic mutations or have developed sporadic ovarian cancer via non-BRCA1/2-associated pathways. These hypotheses are not fully explored in this study. Although mutational signatures appear to be a prognostic indicator that is independent of BRCA1/2 mutations, the overlap and redundancy between mutational signatures and other genomic features associated with BRCA1/2 mutation, such as genomic instability and loss of heterozygosity, are not characterized.

Despite these limitations, our findings support the concept of BRCAness in a subset of ovarian high-grade serous carcinomas without BRCA1/2 mutation, as defined by a signature of somatic events in the cancer genome. We liken the analysis of mutational signatures to that of gene expression profiles, a set of data rooted in biology with a potential to affect clinical decision making. We hypothesize that the improved survival of patients with a BRCA1/2 mutational signature is driven by enhanced sensitivity to platinum-based chemotherapy, and these patients may be candidates for PARP inhibitor therapy. These hypotheses would need to be tested in controlled trials.

This analysis demonstrates the utility of mutational signatures to connect tumorigenic processes, molecular phenotype, and outcome and shows that whole exome sequencing can provide useful prognostic information derived from mutational patterns beyond what can be obtained from more targeted sequencing panels or single gene analysis. In ovarian high-grade serous carcinoma, a BRCA1/2 mutational signature predicts improved overall survival independent of BRCA1/2 mutational status alone. These findings suggest a potential indication for the broad sequencing of ovarian cancer genomes to inform prognosis.

No potential conflicts of interest were disclosed.

Conception and design: F. Dong

Development of methodology: F. Dong, A.H. Beck

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): F. Dong, P.K. Davineni, B.E. Howitt, A.H. Beck

Writing, review, and/or revision of the manuscript: F. Dong, B.E. Howitt, A.H. Beck

Study supervision: F. Dong

The authors gratefully acknowledge members of the TCGA Research Network and the Australian Ovarian Cancer Study for their role in the generation of data used in this study and for their promotion of open access data sharing within the scientific community.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.