Structural Variants at the BRCA1/2 Loci are a Common Source of Homologous Repair Deficiency in High-grade Serous Ovarian Carcinoma

Homologous recombination deficiency (HRD) is identifiable in many cancers and is particularly prominent in high-grade serous ovarian cancer (HGSOC; ref. 1), affecting around 50% of tumors (2) and leaving detectable mutational spectra across the tumor genome (3). The mutational landscape of HGSOC is dominated by extensive genomic copy-number changes and structural rearrangement driven by chromosome instability and defective DNA repair, rather than the patterns of recurrent point mutation in tumor suppressor and oncogenes often observed in other solid tumors (4, 5).

Germline short mutations (GSM) disrupting the coding sequence of BRCA1 and BRCA2 are the most common types of HRD-associated defect, occurring in 8% and 6% of patients with HGSOC, respectively, while disruptive somatic short mutations (SSM) in these genes are present in an additional 4% and 3% of patients with HGSOC, respectively (6, 7). These GSMs and SSMs include single-nucleotide variants (SNV), as well as short indels, with frameshifts being the predominant mechanism of inactivation. These BRCA-deficient tumors represent approximately 20% of patients with HGSOC. An additional 11% of tumors are thought to be BRCA deficient through epigenetic silencing of BRCA1 (2, 8). Mutational or epigenetic inactivation of other genes involved in the homologous recombination (HR) pathway is also thought to confer HRD in a smaller proportion of patients with HGSOC (7, 9–12). Genome-wide patterns of SNVs, indels, and structural variation have been identified as strong predictors of BRCA1/2 deficiency (3, 13). These mutational signatures of BRCA1/2 deficiency are also found in additional tumors that lack short variants at BRCA1/2, suggesting that other unknown aberrations may also be involved in HRD (3). The demonstration of BRCA1/2 loss and detection of HRD are crucial in the management of high-grade serous ovarian (14–19), breast (20, 21), pancreatic (22), and prostate (23) cancer to identify patients whose outcome is markedly improved by the administration of PARP inhibitors (14–16). PARP inhibitors selectively kill HRD cells because these cells are deficient in HR repair (HRR) and can neither resolve stalled replication forks nor accurately repair the increased number of double-strand breaks that result from the use of these agents (24).

The clinical importance of GSMs and SSMs at BRCA1/2 is well established in cancer (17, 20, 25–28). In contrast, the abundance and effects of structural variants (SV) at BRCA1/2 are not well understood, particularly for large SVs (>1 Mb) encompassing multi-megabase regions. Similarly, the compound effects of SVs and short mutations occurring simultaneously at BRCA1 and BRCA2 are poorly studied. Matched tumor-normal whole-genome sequencing (WGS) of freshly frozen tissue is accepted as the best resource to accurately detect SVs in tumors, but in the past such data have been scarce for HGSOC (29, 30). Here, we comprehensively characterize the mutational landscape of BRCA1/2 in HGSOC using the largest collection to date of uniformly processed WGS data (N = 205), comprising two previously published cohorts (5, 6), as well as a large novel cohort described here for the first time. We document the prevalence of HRD across these three cohorts to reveal the complexity of the mutations associated with HRD, their impact on gene expression, and associations with clinical outcome.

WGS data from matched tumor and normal blood samples were uniformly remapped and analyzed to generate a range of somatic mutation calls (Supplementary Fig. S1) for three HGSOC cohorts: chemoresistant or relapsed tumors from the Australian Ovarian Cancer Study (AOCS; ref. 5; N = 80), pretreatment WGS primary tumors from The Cancer Genome Atlas (TCGA; ref. 6; N = 44), and the previously unpublished primary tumors from Scottish High-grade Serous Ovarian Cancer (SHGSOC) study (N = 81) of which 16 (20%) have had neoadjuvant chemotherapy. The combined cohort (N = 205) presented here was uniformly analyzed as follows.

Scottish HGSOC samples were collected via local bioresource facilities at Edinburgh, Glasgow, Dundee, and Aberdeen and stored in liquid nitrogen until required. Patients with HGSOC were determined from pathology records and were included in the study where there was matched tumor and whole-blood samples. Tumor samples were divided into two for DNA and RNA extraction and slivers of tissue were taken, fixed in formalin, and embedded in paraffin wax (FFPE). Samples were only included if they were confirmed as HGSOC and there was greater than 40% tumor cellularity throughout the tumor, determined by using hematoxylin and eosin staining of the FFPE sections and pathology review. Somatic DNA was extracted from the tumor and germline DNA was extracted from whole blood. Quality control was then carried out on the resultant DNA to ensure sufficient purity and quality (Supplementary Materials and Methods). Only when all quality control requirements were satisfied, the DNA was sequenced at the Glasgow Precision Oncology Laboratories.

This study was carried out in accordance with the principles of the Declaration of Helsinki. Ethical approval for the use of human tissue specimens for research was obtained from the Lothian NHS Research Scotland Human Annotated Bioresource (ethics committee reference, 15/ES/0094-SR494). Correlation of molecular data to clinical outcome and clinicopathologic variables in ovarian cancer was approved by South East Scotland Research Ethics Committee 2 (reference, 2007/W/ON/29). All relevant ethical regulations were complied with, including the need for written informed consent where required.

WGS and RNA sequencing (RNA-seq) reads were downloaded in compressed FASTQ format from the sequencing facility (SHGSOC) or in aligned BAM format (including unaligned reads) from the European Genome/Phenome Archive (AOCS) and the Bionimbus Protected Data Cloud (TCGA_US_OV). The reads obtained in BAM format were query sorted and converted to FASTQ. All whole-genome sequence and RNA-seq data for the SHGSOC cohort will be made available on publication via European Genome-Phenome Archive (EGA).

Reads were aligned to the GRCh38 reference genome and somatic and germline variant calling was run using a bcbio (31) 1.0.7 pipeline (see Supplementary Data for full pipeline configuration, program, resource versions, and references). Germline SNPs and indels were called with GATK 4.0.0.0 HaplotypeCaller. Somatic SNVs and indels were called as a majority vote between Mutect2, Strelka2, and VarDict. Small variants were annotated with Ensembl Variant Effect Predictor v91 and filtered for oxidation artefacts by GATK 4.0.0.0 FilterByOrientationBias. Somatic SVs were called with Manta 1.2.1 and somatic copy-number variants with CNVkit 0.9.2a0. LOH and somatic copy-number variants were also identified with CLImAT. Sample quality control was performed with Qsignature 0.1 to identify sample mix-ups and with VerifyBamId 1.1.3 to identify sample contamination. Tumor cellularity was estimated using both CLImAT's estimates and p53 variant allele frequency. These measures were compared with the qPure estimates for the AOCS cohort (5) and histopathologic estimates for the SHGSOC cohort with very good concordance (Supplementary Fig. S2). The CLImAT estimates were used as the final estimates of cellularity.

Germline short variants at BRCA1/2 were filtered to include only damaging pathogenic variants for the purposes of establishing BRCA1/2 mutational status. Included variants were all of moderate or high impact according to variant effect predictor (VEP). Variants with a pathogenic or risk factor annotation according to ClinVar (32) were included (n = 145). Remaining variants with a ClinVar benign or likely benign status were excluded (n = 1,147). Remaining frameshift or nonsense (stop gained) or splice donor/acceptor variants were included (n = 125). Remaining missense variants with damaging SIFT and PolyPhen predictions were included (n = 36). Remaining missense variants called as damaging by only one of SIFT and PolyPhen were considered borderline and were excluded if their CADD score < 20 (33). Missense or in-frame variants with no Clinvar, SIFT, or PolyPhen evidence were excluded.

Somatic short variants at BRCA1/2 were also filtered for pathogenicity to include variants that: were annotated by VEP as being of high or moderate impact, were pathogenic according to at least one of SIFT or PolyPhen, and had a high CADD score. In addition, we excluded somatic variants with an allele frequency less than 0.4.

We identified SVs in patients with HGSOC using Illumina paired and split read–based SV detection tool, Manta (34). However, we observed that Manta was failing to detect a large number of very large deletions (>1 Mb) that had been identified using depth of coverage–based approaches in the Pan-Cancer Analysis of Whole Genomes (PCAWG; ref. 35). Therefore, we chose to supplement these calls with copy-number variants greater than 1 Mb in size that were called by one caller (CNVkit) using evidence from read depth and were also confirmed by an allele-specific copy-number caller (CLImAT), which provides an additional layer of evidence in addition to read depth as it also incorporates the shift in allele frequency of heterozygous SNPs within the potential copy-number variant into its variant calling algorithm, which also accounts for aneuploidy and sample cellularity. Deletions were assumed to be heterozygous if the copy number as estimated by both CNVkit and CLImAT was 1, although this may be a conservative estimate of allelic loss in the presence of subclones. We visually inspected all the identified SVs in the Integrative Genomics Viewer (v2.4.10). The magnitude of coverage and log fold change were inspected to confirm either duplication or deletion. For SVs in the range 300 bp–30 kb, the paired-end sequencing reads were manually reviewed, including looking at split reads, paired-end insert size, read coverage, and pair-orientation. We compared our filtered set of large copy-number variations (CNV) in the samples included in PCAWG with PCAWG's copy-number calls and found that 40 of 41 of our variants in these samples were also identified by PCAWG.

Permutation analyses were carried out using the R package RegioneR (36) to investigate whether large deletions overlap more often with BRCA1 and BRCA2 than they do elsewhere in the genome. We carried out 100,000 permutations to simulate the null hypothesis throughout the genome for each gene and judged significance at alpha = 0.05 (Supplementary Fig. S3).

We investigated whether large deletions, large duplications, and inversions are enriched at BRCA1 and BRCA2 within their respective chromosomes using 100,000 circularized permutations to generate the null distribution of overlaps in each case (Supplementary Fig. S4). The observed number of overlaps with BRCA1/2 was well within the range of this null distribution and, therefore, showed no evidence of within-chromosome enrichment, which is perhaps unsurprising given the large sizes of these events relative to the length of their chromosomes.

To predict the level of HR deficiency in each tumor sample, we implemented the HRDetect algorithm as published by Davies and colleagues (3). We based our implementation on a Snakemake pipeline made publicly available by Zhao and colleagues (37), with some modifications to ensure accurate recapitulation of the original method (Supplementary Materials and Methods). As some of the AOCS cohorts' patients included here were also used in the validation of HRDetect in the original publication, we were able to compare our implementation for the same patients with that of the authors. Our implementation of HRDetect was very highly correlated with the original HRDetect implementation on the same samples (Spearman rho = 0.92).

We used the weights for the independent variables that were defined by the original model rather than retraining the model on our data as the original weights trained on a breast cancer dataset have been shown to perform well on ovarian cancer datasets and our total sample size was substantially smaller than that used to train the model originally. As input, we used the somatic SNVs and indel calls identified by the ensemble calling approach described above, with the SV calls made by Manta and the copy-number segments defined by CNVkit.

HGSOC samples were collected and underwent quality control as described for the DNA samples used for WGS. Somatic RNA was extracted from the resulting RNA sample, underwent quality control, and was quantified. RNA-seq was carried out by the Edinburgh Clinical Research Facility (Edinburgh, Scotland, United Kingdom) on an Illumina NExtSeq500 (further details in Supplementary Materials and Methods).

RNA-seq data were analyzed using the Illumina RNA-seq best practice template. Briefly, reads were aligned to the GRCh38 reference genome and quality control was carried out. Salmon quant (38) was used to quantify the expression of transcripts against the GRCh38 RefSeq transcript database indexed using the salmon index (k-mers of length 31). Transcript-level abundance estimates were imported into R and summarized for further gene-level analyses. For differential expression analyses, raw expression counts were used by the DESeq2 package (39). For visualization of gene expression, counts were normalized using the variance stabilizing transformation. Previously published RNA-seq data available for the AOCS (ref. 5; N = 80) and TCGA (ref. 6; N = 30) cohorts, together with novel RNA-seq data for the SHGSOC (N = 40) cohort, generated for this study as detailed above, were processed in this way from FASTQ.

Clinical data for the SHGSOC cohort were retrieved from the Edinburgh Ovarian Cancer Database (40), the CRUK Clinical Trials Unit Glasgow, and available electronic health records (ethics reference, 15/ES/0094-SR751).

The clinical information, including survival endpoints, age, and stage at diagnosis, is available for these patients as part of the PCAWG project (35).

All downstream statistical analyses were carried out in R (v3.6.0) using Jupyter notebook (v4.3.1).

To compare the gene expression levels at BRCA1/2 between tumors with and without BRCA1/2 deletions, we used the package DESeq2 to test for differential expression between the raw gene expression counts at each gene between samples with and without a deletion at that gene. At BRCA1, samples that also had a short variant had significantly lower expression than those that had a deletion alone. As a result, we only considered the samples with a deletion in the absence of a short variant. At BRCA2, this was not the case and the samples with short variants in addition to a deletion had comparable levels of expression with those with deletions alone, and were therefore included in the analysis. Cohort and tumor sample cellularity were included as covariates in the model formula.

The risk of HR deficiency in tumors with BRCA1/2 mutations, grouped by type, relative to those tumors without BRCA1/2 mutations, was determined by using Fisher exact tests. The effect of mutations at BRCA1 and BRCA2 was determined together and, where sample size permitted, separately for mutational categories including: GSMs only, SSMs only, single deletion which overlaps at least one exon in the absence of a short variant or other SVs, deletion of at least one exon at both BRCA1 and BRCA2 in the absence of a short variant, inversion spanning BRCA1 in the absence of short variants, or deletion and duplication spanning BRCA2 in the absence of short variants or deletion. We also considered the impact of the presence of a short variant accompanied by a deletion at either of the genes. These categories are further described together with their labels and color coding in the Supplementary Data. The RR conferred by each mutational category was calculated in comparison with the group of patients without BRCA1/2 mutations or SVs. Samples where BRCA1/2 promoter methylation had been detected were excluded, except for where the effect of BRCA1 promoter methylation was itself being examined. All samples with BRCA1 promoter methylation are predicted to be HR deficient so pseudocounts of 1 are used to estimate the effect size which is, therefore, a likely underestimate. P values were adjusted for the impact of multiple testing using Benjamini–Hochberg correction and were considered together with the effect sizes in the reporting of results.

Given the relative sparsity of the data and the correlation between features, we used a multivariable elastic net regularized regression model for the binary outcome of HRD defined by a probability of HRD greater than 0.7 from HRDetect. The data were partitioned into train and test sets (80:20) and the tuning parameters were optimized, to maximize the AUC, using 10-fold cross-validation of the training set. The model was then fitted to the training set and the model performance was assessed using the test set. The input variables available for selection were: BRCA1 germline short variant status, BRCA1 somatic short variant status, the presence of a large somatic deletion at BRCA1, the presence of a large somatic deletion at BRCA1 and a BRCA1 short variant, all the corresponding variables for BRCA2, the presence of an inversion at BRCA1, the presence of a duplication at BRCA2, the presence of a large somatic deletion at BRCA1 and at BRCA2 (double deletion), BRCA1 promoter hypermethylation, whole-genome doubling, genome-wide load of SNVs, large CNVs, and SVs, in addition to cohort and tumor cellularity. The data were partitioned and the model was optimized and fitted to 100 train-test splits of the data to assess the robustness of the feature selection.

Follow-up information, including overall survival time, was available for 190 of 205 patients, of which 144 were deceased by the time of last follow-up. The association between genome-wide patterns of HRD and progression-free survival (PFS) was also assessed. Progression-free interval time was available for 151 of the patients from the AOCS and SHGSOC cohorts, of which 129 relapsed by the time of last follow-up. The effect of the HRDetect score, as a measure of the probability of HRD in the tumor, on the length of time that patients survived after diagnosis (overall survival time) and the time between diagnosis and first radiologically defined progression (PFS time) was assessed using Cox proportional hazards models stratified by cohort. Multivariable models were also fitted adjusting for age and stage at diagnosis and the Schoenfeld residuals were examined. In all cases, hazard ratios reported corresponded to a one SD increase in HRDetect score. Survival probability through time was compared between HR-deficient (HRDetect score > 0.7) and HR-proficient (HRDetect score ≤ 0.7) patients in Kaplan–Meier plots. This was repeated excluding the patients with BRCA1/2 short variants to assess the impact of HRD driven by other events on survival.

We used the consensus SV and CNV calls generated and included in the 2017-01-19 release of PCAWG to investigate the abundance of SVs at BRCA1/2 in other cancer types. Samples (N = 2,567) were from all cancer subtypes included in PCAWG that are recommended for use in subtype-specific analyses. Disruptive SVs were identified by intersecting the variant calls with the exons of BRCA1/2. Abundance of SVs, deletions, and duplications were tabulated separately for each of the cancer types. Cancer subtypes were ordered from left to right by overall SV burden as determined by PCAWG.

Using gene expression quantifications from matched RNA-seq for N = 735 of these samples, we compared BRCA1 expression in samples with and without BRCA1 deletions from each cancer type separately in differential expression analyses carried out using DESeq2 (39). Cancer types were included if there were greater than 15 samples in total and if at least three of these samples harbored deletions. This effect was also measured across all samples pan-cancer adjusting for differences in expression between cancer types in the model. These analyses were repeated for BRCA2 expression. Effects on expression and the data behind them were displayed after variance stabilizing transformation of the expression counts.

Previously published WGS and RNA-seq data that were reanalyzed here are available via EGA at accession code EGAS00001001692 (ICGC PCAWG). WGS, RNA-seq, and clinical data from the Scottish cohort (SHGSOC) will be made available via EGA at accession code EGAS00001004410. Other supporting data have been provided in the Supplementary Tables.

All codes will be made available at: https://github.com/ailithewing/Structural_variants_BRCA1_2_HRD_inHGSOC.

A variety of SVs were detected at the BRCA1/2 loci, but large multi-megabase deletions spanning the entirety of BRCA1 or BRCA2 dominated. In all three cohorts, some deletions encompassed more than 10% of chromosomes 17 or 13, although the majority were more focal (median BRCA1 deletion = 4.9 Mb and median BRCA2 deletion = 6.2 Mb; Fig. 1). Heterozygous deletions occurred at similar rates at BRCA1 (16%) and BRCA2 (14%) overall, and at comparable rates between the similarly sized AOCS and SHGSOC cohorts (Supplementary Table S1). In six of 205 (3%) samples in the combined cohort, we observed large deletions at both BRCA1 and BRCA2 in the absence of short mutations at both genes. Permutation analyses revealed that BRCA1 and BRCA2 were deleted more than expected given the observed distribution of large deletions throughout the HGSOC genome (Supplementary Fig. S3). This is consistent with selection for these events as has been reported for SNVs at BRCA1/2 (41), but we cannot exclude the role of mutational bias in producing these enrichments of deletions. Inversions occurred less often than deletions, but did occur in isolation in 6% of samples, and within groups of large overlapping inversions in 5% of samples. In addition, we observed large duplications that spanned the entire length of either BRCA1 or BRCA2 in all cohorts (2% and 7%, respectively; Supplementary Table S2). LOH was near ubiquitous at BRCA1 (202/205; 99%), of which 166 events were copy-number neutral, and was present at BRCA2 in more than half of samples (118/205; 58%), of which 91 events were copy-number neutral. We observed similar genome-wide SV mutational spectra in each cohort despite the clinical differences among them: in particular AOCS represents chemoresistant/relapsed cases, while SHGSOC is composed mainly of samples taken before treatment. This suggests that large SVs predicted to impair BRCA1/2 function are a general feature of HGSOC evolution.

We found that patients with a large deletion spanning BRCA1, in the absence of a GSM or SSM, had lower BRCA1 expression than those patients with no BRCA1 GSM, SSM, or SVs (log₂ fold change of no GSM/SSM/SVs vs. a deletion only = 0.45; P = 0.0093; Fig. 2). We also found that tumors with a large deletion spanning BRCA2 had lower BRCA2 expression than those without GSM/SSM/SVs at BRCA2 (log₂ fold change in expression between no GSM/SSM/SVs vs. a deletion = 0.43; P = 0.037). In the analysis of BRCA2 expression, we included tumors with large deletions that also had GSM or SSM in the remaining copy, having shown that in our data, BRCA2 expression is not significantly different in these tumors (Supplementary Fig. S5). In spite of the unavoidable heterogeneity in tumor expression data among samples, although variation in tumor purity was accounted for, the trends observed here are consistent with large BRCA1/2 deletions, reducing BRCA1/2 expression, although indirect mechanisms cannot be excluded.

We examined the functional impact of all BRCA1/2 mutations detected across all cohorts using an established method, HRDetect (3), which predicts HRD based upon genome-wide mutational spectra, and, therefore, provides a functional readout for the HRR pathway in tumors (Fig. 3).

As expected, tumors with GSMs or SSMs in BRCA1/2 are more likely to have HR-deficient tumors (HRDetect scores > 0.7) than those tumors without GSMs, SSMs, or SVs at these genes [GSM OR, 6.9; 95% confidence interval (CI), 1.8–33; P = 2.3 × 10⁻³; P_adj = 3 × 10⁻² and SSM OR, 25; 95% CI, 3.2–1,121; P = 1.7 × 10⁻⁴; P_adj = 2.3 × 10⁻³; Fig. 3). Four samples with GSMs at BRCA2 demonstrated low HRDetect scores and accordingly showed no evidence for subsequent loss of the wild-type allele in the tumor which suggests that certain GSMs that are predicted to be disruptive are insufficient to generate HRD (Supplementary Table 3a).

The majority of samples (65%) with BRCA1/2 deletions were HRD (HRDetect score > 0.7), although some deletions occurred in tumors that also harbored short mutations (19/49). Because many of the samples with BRCA1/2 deletions lacked short mutations (30/49; 61%), analysis of the effects of deletions independent of short variants was undertaken. The compound effect of deletions at both the BRCA1 and BRCA2 loci in the absence of short variants was particularly pronounced, demonstrating a significantly increased risk of HRD (OR, 19; CI, 2.4–896; P = 1.3 × 10⁻³; P_adj = 1.7 × 10⁻²) and consistently high HRDetect scores. Furthermore, compound deletions at both loci generated an OR that was comparable with other classes of HRD mutations known to have clinical importance, including the well-studied disruptive BRCA1/2 short variants (Fig. 3C and D). Single deletions at either BRCA1 or BRCA2 did not consistently confer an increased risk of HRD in the absence of a BRCA1/2 short variant (Fig. 3), although the analysis may be underpowered to detect small effects given the current sample size. The HRDetect scores for samples with these single deletions formed a bimodal distribution for which we have been unable to find a defining characteristic for the difference, such as the length of the deletion, resultant level of gene expression, or a background of whole-genome doubling. The estimated effect that we observe of a single deletion at BRCA2 merits further investigation, although it did not achieve statistical significance in our data (OR, 2; 95% CI, 0.37–10.3). Previous studies have identified rearrangement signatures associated with HR deficiency (42). We observed an elevation in these signatures, particularly in the case of rearrangement signature 5, in samples with deletion at BRCA1/2, which is broadly consistent with the levels of elevation observed in the presence of short mutations at BRCA1/2 (Supplementary Fig. S6). However, HRDetect incorporates additional genome-wide sources of information, including the more powerful presence of indel microhomology leading to more accurate HRD prediction.

Beyond deletions, the functional impact of other classes of SV, such as inversions or duplications, is less well studied. Half of the samples with only BRCA1 inversions bear an HR-deficient signature, which suggests that although in isolation their presence is not associated with HR deficiency, these samples can show evidence of HRD perhaps via another unappreciated route (Fig. 3). In contrast, only one of the samples with only BRCA2 duplications was HR deficient, which suggests potential for enrichment in HR-proficient samples, but this would need to be further explored in greater sample sizes.

Samples across the combined cohort never had more than one short mutation across BRCA1 and BRCA2. This suggests that SSMs are not a mechanism for biallelic inactivation of a gene affected by a GSM. This is consistent with reports from a previous study of HGSOC (ref. 41; Supplementary Table S3A and S3B). In contrast, of the HGSOC tumors with a GSM at BRCA1/2 predicted to cause HRD, we found that 11 of 32 (34%) showed evidence for an SV at the same gene which may contribute to HRD if the SV occurs on the other allele (Supplementary Tables S2 and S3). We found that most of these somatic events (8/11; 73%) were large deletions, while two tumors possessed more than one SV spanning the same gene as the GSM and one more showed evidence of somatic duplication. The importance of “second hit” (43) mutations in tumors is well established (44), but these data suggest that multi-megabase deletions have an underappreciated role in this phenomenon in HGSOC. Across the three cohorts, 24% (50/205) of patients had a disruptive short variant at either BRCA1 or BRCA2, and 30% (15/50) of these patients also carried a BRCA1/2 deletion at the same locus. Also, it appears that SVs, including deletions, can occur at both BRCA1 and BRCA2 in the same sample. Large somatic deletions occurred at both BRCA1 and BRCA2 in 13 samples and in seven of these samples, there was no short variant at either gene, although one sample had a hypermethylated BRCA1 promoter. These data suggest that large deletions and other SVs disproportionately contribute to biallelic inactivation in HGSOC, driven by the unusually high rates of structural variation seen in this cancer (45).

We comprehensively modeled the effects of a range of genomic alterations at the BRCA1/2 loci on HRD, to investigate the relative importance of these features in explaining the patterns of HRD observed. Given the relative sparsity of the data and the correlation between features, we used a multivariable elastic net regularized regression model.

In addition to the previously reported impact of short variants at BRCA1/2 and BRCA1 promoter hypermethylation on HRD, large deletions at BRCA2 conferred an increased risk of HRD (Fig. 4). Furthermore, samples with double deletions, where deletions were found at both BRCA1 and BRCA2, were more likely to be HR deficient. Importantly, the influence of these double deletions on HRD exceeded that of genome-wide large CNV loads and genome-wide SV loads. Also, large inversions at BRCA1 were independently associated with an increased risk of HRD. The functional impact of these events on the gene is currently unknown, but this suggests that these events may either be markers for processes that impact the gene's function or may even directly impact the function of the gene themselves. The model's ability to predict HRD was good, with a mean ROC curve AUC of 0.75, which although promising, suggests that there are additional unknown sources of HRD.

We can explain the observed pattern of HRD by the presence of mutational or epigenetic defects at the BRCA1/2 genes in 81 of 106 samples with predicted HRD (72 GSM/SSM/SVs at BRCA1/2, nine with BRCA1 promoter methylation), but a further 25 samples with HRD remain unexplained. On further examination, we found that all of these samples harbored damaging GSMs and/or SSMs at other HR genes (defined by Kyoto Encyclopedia of Genes and Genomes pathway annotation; Supplementary Tables S5–S7), motivating analysis of the potential roles of mutations at loci other than BRCA1/2 and their inclusion in an expanded model. However, we found no convincing evidence for a strong influence of mutations at other HR genes or the combined expression of genes dysregulated in the presence of HRD (Supplementary Fig. S7).

In HGSOC, while HRD as a result of GSMs and SSMs at BRCA1/2 is associated with response to platinum and PARP inhibition and improved survival, the relationship between HRD resulting from disruption of BRCA1/2 via other mechanisms is less clear (1, 6, 8, 11, 46–49).

A higher probability of HRD was significantly associated with longer overall survival in our combined cohort [Hazard Ratio, 0.64; 95% CI, 0.53–0.76 (per 1 SD increase); P = 8.4 × 10⁻⁷] and this effect was only slightly attenuated by adjustment for patient age and tumor stage at diagnosis (Fig. 5A). Notably, this effect persisted [Hazard Ratio, 0.67; 95% CI, 0.56–0.84 (per 1 SD increase); P = 3.6 × 10⁻⁴; Fig. 5B] when we excluded the patients with BRCA1/2 GSM/SSM, who are already known to have longer survival. We saw a similar effect when we examined the effect of HRD on PFS [Hazard Ratio, 0.77; 95% CI, 0.64–0.93 (per 1 SD increase); P = 0.007], which also was robust to the exclusion of patients with BRCA1/2 GSM/SSM [Hazard Ratio, 0.79; 95% CI, 0.64–0.98 (per 1 SD increase); P = 0.03]. Correspondingly, in all of these instances, patients with HRD tumors (HRDetect > 0.7) had longer survival times than patients with non-HRD tumors. The effects observed are consistent with the presence of HRD through mechanisms other than BRCA1/2 GSMs and SSMs affecting overall survival.

Cancers with other sites of origin vary in their level of structural variation. It is possible that large deletions may compromise BRCA1/2 function in these cancers also. The PCAWG consortium recently generated conservative consensus SV calls based upon WGS data across many tumor types, but did not report specifically on SV patterns at BRCA1/2 (50). These data illustrate that large deletions (Fig. 6A; and other SVs; Supplementary Fig. S8) at BRCA1/2 are common across a variety of non-HGSOC cancer types, including in cancers also known to show evidence of HRD, such as breast and prostate cancer. In addition, two rare tumor types showed notably higher frequencies of BRCA1/2 deletions: chromophobe renal cell carcinoma (CHRCC) and leiomyosarcoma. Exploiting matched RNA-seq for N = 735 of the PCAWG samples, we found that BRCA1 and BRCA2 have significantly lower expression in those samples with deletions at BRCA1 and BRCA2, respectively, in comparison with those samples without a deletion (log₂ fold change of no BRCA1 deletion vs. a BRCA1 deletion = 0.94; P = 5.9 × 10⁻¹² and log₂ fold change of no BRCA2 deletion vs. a BRCA2 deletion = 0.6; P = 2.9 × 10⁻⁶) after adjusting for primary site. Furthermore, BRCA1 expression was consistently compromised within all cancer types with sufficient samples to test individually (Fig. 6B and C).

DNA repair deficiency in general and HRD in particular represent cancer cell vulnerabilities that have recently been exploited to exceptional patient benefit (14, 15, 17, 19–24, 46). However, the understanding of the genomic mechanisms that give rise to HRD is incomplete and the identification of patients whose tumors are HR deficient remains inaccurate.

In the largest collection of HGSOC WGS data examined to date, with matched expression data for most of the 205 tumors included, we have revealed new insights into the genesis of HRD in HGSOC based upon genome-wide mutational spectra. We show that structural variation at BRCA1/2 in HGSOC is frequent and dominated by multi-megabase deletions, which are significantly enriched at BRCA1 and BRCA2 relative to the rest of the HGSOC genome. Examining transcriptomic data for the same samples, we have shown that large deletions overlapping BRCA1/2 are associated with lower BRCA1/2 expression, suggesting a direct impact of these deletions on gene function in many cases. Large deletions spanning BRCA1/2 contribute to HRD independently of short variants, and samples with compound deletions affecting both BRCA1 and BRCA2 generate the highest risk of HRD. The frequent inactivation of BRCA1/2 by large deletions in HGSOC is novel to our knowledge, and the original analysis of the AOCS cohort reported only one BRCA1/2 large deletion (>1 Mb; ref. 5; Supplementary Table S4.2). The original TCGA cohort analysis did report frequent losses of the chr13q and chr17q chromosome arms (including the BRCA1/2 loci) based upon SNP microarray data (ref. 6; Supplementary Table S5.1), but such data are known to generate high levels of false positives and negatives (47, 48) and these losses were not postulated to affect BRCA1/2 function. Thus, previous assessments of SVs impacting the BRCA1/2 loci have been characterized by underreporting, likely to be a result of the use of less sensitive algorithms tuned to detect smaller focal deletions (5), as well as copy-number alteration estimates derived from SNP microarray data (5, 6) and exome-restricted sequencing data (6, 49). Other types of structural variations are less frequent, but still evident, such as large inversions at BRCA1 and duplications at BRCA2. The impact of these categories of mutation on the function of the gene is less well studied, but our data suggest that when BRCA1 inversions in particular are considered together with the other mutational events in the tumor, their presence may aid prediction of HRD.

Our data also suggest a significant frequency of BRCA1 and BRCA2 loss by structural variation in a series of other cancers, including both cancers known to suffer BRCA1 and BRCA2 loss due to SNVs or indels, such as breast, pancreatic, and prostate cancer, as well as a number of cancers outside of this group, including squamous cell lung cancer and cervical cancer. Perhaps most notably, soft-tissue leiomyosarcoma had a high incidence of large deletions at BRCA2 (47%) and chromophobe renal cancers had a high incidence of large deletions of both BRCA1 (53%) and BRCA2 (47%). CHRCC often undergoes concurrent deletion of both BRCA1 and BRCA2 as a result of whole chromosome arm losses as has been reported previously (51), and trials of PARP inhibition as a therapy for renal cell carcinoma are ongoing (52). The fact that across these cancer types, large BRCA1 and BRCA2 deletions were associated with loss of gene expression suggests functionality is lost and that strategies to detect these genomic events and trials of PARP inhibition should also be considered in these patients.

Finally, we have constructed an integrated model of HRD in HGSOC, including a large variety of mutation- and expression-based variables across the combined cohort. This model supports an independent role for structural variation at BRCA1/2 in HRD and highlights the diversity of routes that tumors may follow to reach HRD. Given this diversity, and the substantial fraction of samples where HRD is detected in the absence of any detectable BRCA1/2 mutations, we conclude that the direct detection of HRD in HGSOC using genome-wide sequencing data is a valuable addition to the search for inactivating mutations in HR pathway genes. This is likely to be the case for other cancers showing evidence for HRD, such as uterine, lung squamous, esophageal, sarcoma, bladder, lung adenocarcinoma, head and neck, and gastric carcinomas (1). The variety of events sufficient for a tumor to develop HRD is not well understood, but recent studies suggest that there is selective pressure for biallelic inactivation leading to HRD in cancer types with predisposing germline variants in the HR pathway, such as breast, ovarian, pancreatic, and prostate cancers (53).

One of the key challenges in studies of this type is deciding upon a “gold standard” test of HRD. Current functional, clinical, and molecular tests all have advantages and disadvantages. The limitations of HRDetect that we used here include that it was developed and trained using breast tumor data and is predicated on BRCA1/2 deficiency arising from short variant disruption and promoter methylation, rather than any form of disruption to any HR gene. Although in the context of this study, of BRCA1/2 disruption by SV, the latter is of less importance. All current genomic HRD tests are further limited to demonstrating that HRD once existed in the evolution of a tumor, and are blind to the restoration of HR by events such as secondary mutations, hypomorphic HRD variants, and epigenomic changes.

There is an urgent clinical need to better understand the processes that give rise to both BRCA1/2 loss and more broadly contribute to HRD. Our study demonstrates that BRCA1/2 loss by structural variation may have a comparable impact on HRD and patient survival with short variants at BRCA1/2. Furthermore, these events are not specific to HGSOC, they are abundant, they compromise gene expression, and are likely to be functionally important in other cancer types, including in those types where HRD has been identified previously. However, these variants are unlikely to be detected by sequencing methods currently employed in the clinic. A change in sequencing approach to identify all BRCA1/2 loss may be required to maximize the potential of PARP inhibition.

A. Ewing reports grants from UKRI during the conduct of the study. R.L. Hollis reports grants from AstraZeneca during the conduct of the study and personal fees from GlaxoSmithKline outside the submitted work. C.S. Herrington reports grants from AstraZeneca during the conduct of the study. N. McPhail reports personal fees from Eisai, Ipsen, and Bayer outside the submitted work. S. Dowson reports grants from Beatson Cancer Charity during the conduct of the study. R. Glasspool reports grants from AstraZeneca during the conduct of the study, as well as grants from Clovis Oncology and personal fees from AstraZeneca, Clovis Oncology, and GlaxoSmithKline outside the submitted work. B. McDade reports grants from Scottish Genomes Partnership during the conduct of the study. A. Matakidou reports other from AstraZeneca during the conduct of the study and GlaxoSmithKline outside the submitted work, and has AstraZeneca shares. B. Dougherty reports other from AstraZeneca during the conduct of the study and outside the submitted work. R. March reports employment with and is a shareholder of AstraZeneca; AstraZeneca is a pharmaceutical company, with commercial products in PARP inhibition. J.C. Barrett reports other from AstraZeneca during the conduct of the study and outside the submitted work. I.A. McNeish reports personal fees from Clovis Oncology, GlaxoSmithKline, Roche, Alkerrnes, and ScanCell and grants and personal fees from AstraZeneca outside the submitted work. A.V. Biankin reports other from Bristol Myers Squibb, AstraZeneca, MyTomorrows, Elstar Therapeutics, Cumulus Oncology, Modulus Oncology, Wollemia Oncology, ConcR, Cambridge Cancer Genomics, Agilent Technologies, Gabriel Precision Oncology, and Human.ai outside the submitted work. P. Roxburgh reports grants from AstraZeneca during the conduct of the study, as well as grants from AstraZeneca and personal fees from AstraZeneca and GlaxoSmithKline/Tesaro outside the submitted work. C. Gourley reports grants from AstraZeneca and Scottish Chief Scientist's Office during the conduct of the study. C. Gourley also reports grants from Novartis, Aprea, BerGenBio, and Medannexin; grants and personal fees from AstraZeneca, GlaxoSmithKline, Tesaro, MSD, Roche, and NuCana; and personal fees from Clovis, Foundation One, Chugai, Takeda, Sierra Oncology, and Cor2Ed outside the submitted work, as well as has a patent for PCT/US2012/040805 issued, PCT/GB2013/053202 pending, 1409479.1 pending, 1409476.7 pending, and 1409478.3 pending. No disclosures were reported by the other authors.

A. Ewing: Conceptualization, data curation, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. A. Meynert: Conceptualization, data curation, formal analysis, investigation, methodology, writing–review and editing. M. Churchman: Resources, data curation, project administration, writing–review and editing. G.R. Grimes: Data curation, formal analysis, investigation, methodology, writing–review and editing. R.L. Hollis: Resources, data curation, writing–review and editing. C.S. Herrington: Conceptualization, resources, data curation, writing–review and editing. T. Rye: Data curation, writing–review and editing. C. Bartos: Data curation, writing–review and editing. I. Croy: Resources, writing–review and editing. M. Ferguson: Conceptualization, resources, data curation, writing–review and editing. M. Lennie: Resources, data curation, writing–review and editing. T. McGoldrick: Conceptualization, resources, data curation, writing–review and editing. N. McPhail: Conceptualization, resources, data curation, writing–review and editing. N. Siddiqui: Resources, writing–review and editing. S. Dowson: Resources, data curation, writing–review and editing. R. Glasspool: Resources, data curation, writing–review and editing. M. Mackean: Data curation, writing–review and editing. F. Nussey: Resources, data curation, writing–review and editing. B. McDade: Resources, writing–review and editing, performed whole genome sequencing. D. Ennis: Resources, writing–review and editing. L. McMahon: Supervision, project administration, writing–review and editing. A. Matakidou: Conceptualization, supervision, writing–review and editing. B. Dougherty: Conceptualization, supervision, writing–review and editing. R. March: Conceptualization, supervision, writing–review and editing. J.C. Barrett: Conceptualization, supervision, writing–review and editing. I.A. McNeish: Conceptualization, supervision, writing–review and editing. The Scottish Genomes Partnership: Conceptualization, funding acquisition, writing–review and editing. A.V. Biankin: Conceptualization, supervision, funding acquisition, writing–review and editing. P. Roxburgh: Conceptualization, resources, data curation, supervision, investigation, writing–review and editing. C. Gourley: Conceptualization, resources, data curation, supervision, investigation, visualization, writing–original draft, writing–review and editing. C.A. Semple: Conceptualization, data curation, supervision, investigation, visualization, writing–original draft, writing–review and editing.

A. Ewing was supported by a UKRI Innovation Fellowship (MR/RO26017/1). C.A. Semple, A. Meynert, and G.R. Grimes were supported by MRC core funding to the MRC Human Genetics Unit (MRC grant MC_UU_00007/16). R.L. Hollis was supported by an MRC-funded Research Fellowship. S. Dowson received funding from AstraZeneca and the Beatson Cancer Charity. I.A. McNeish acknowledges funding from Ovarian Cancer Action and the NIHR Imperial Biomedical Research Centre. A.V. Biankin acknowledges funding from Cancer Research UK (C29717/A17263, C29717/A18484, C596/A18076, C596/A20921, and A23526), Wellcome Trust Senior Investigator Award (103721/Z/14/Z), Pancreatic Cancer UK Future Research Leaders fund (FLF2015_04_Glasgow), MRC/EPSRC Glasgow Molecular Pathology Node, and The Howat Foundation. Sequencing of the SHGSOC cohort was supported by AstraZeneca, the Medical Research Council, and the Scottish Chief Scientist through a Precision Medicine Scotland Innovation Centre/Scottish Genome Partnership (SEHHD-CSO 1175759/2158447) collaboration. This Scottish Genomes Partnership was funded by the Chief Scientist Office of the Scottish Government Health Directorates (SGP/1) and The Medical Research Council Whole Genome Sequencing for Health and Wealth Initiative (MC/PC/15080). This study would not be possible without the families, patients, clinicians, nurses, research scientists, laboratory staff, informaticians, and the wider Scottish Genomes Partnership team to whom we give grateful thanks. Members of the Scottish Genome Partnership (SGP) include Timothy J. Aitman, Andrew V. Biankin, Susanna L. Cooke, Wendy Inglis Humphrey, Sancha Martin, Lynne Mennie, Alison Meynert, Zosia Miedzybrodzka, Fiona Murphy, Craig Nourse, Javier Santoyo-Lopez, Colin A. Semple, and Nicola Williams. More information about SGP can be found at www.scottishgenomespartnership.org. The authors would also like to acknowledge the Edinburgh Clinical Research Facility for the sequencing of RNA samples from the SHGSOC cohort. The authors would also like to extend our thanks to the Nicola Murray Foundation and the Edinburgh Ovarian Cancer Database, from which the clinical data for much of the Scottish cohort were retrieved. We also thank the NRS Lothian Human Annotated Bioresource, NHS Lothian Department of Pathology, the Edinburgh Experimental Cancer Medicine Centre, the Biorepository at the Glasgow Queen Elizabeth University Hospital, and the Tayside Biorepository for their support.

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.