Detection of Molecular Signatures of Homologous Recombination Deficiency in Prostate Cancer with or without BRCA1/2 Mutations

Ovarian and breast cancer are often associated with mutations in BRCA1 and BRCA2, the key enzymes of a specific DNA repair pathway, homologous recombination (HR). Inactivating mutations in these genes often render such tumors homologous recombination deficient (HRD) that lead to sensitivity to PARP inhibitors (PARPi), a novel class of cancer therapy, developed for such tumors based on the principle of synthetic lethality (1). Several PARP inhibitors have been approved over the past several years for the treatment of appropriately selected ovarian and breast cancer cases (2). It was shown that PARP inhibitors are also effective in prostate cancer, which is partly due to the presence of BRCA1 and BRCA2 mutations (3, 4).

HRD can be induced by germline or somatic BRCA1/2 genetic mutations; however, it can also be present in tumors with intact BRCA1/BRCA2 genes. The list of genes involved in homologous recombination is far from complete. For example, loss of CHD1 was recently implicated as a possible mechanism leading to HRD in prostate cancer (5). HRD can also be induced through the suppression of expression of key HR genes by, for example, the inactivation of SPOP in prostate cancer cells (5). Therefore, sequencing panels of HR-related genes will likely miss a subset of HRD cases. Because tumors deficient in HRD are sensitive to PARPi/platinum, the consequence of not identifying tumors that harbor HRD is that treatment with these agents may be delayed or not given. Identification of such tumors is clinically relevant because translational studies indicated that likely HRD tumors without BRCA1 or BRCA2 mutations may also benefit from HRD-associated therapy (6). A particularly promising approach for functionally evaluating HRD focuses on specific DNA aberrations, or “DNA scars” that result from HRD. This approach is based on the principle that the normal function of homologous recombination ensures the error-free repair of double-strand breaks; therefore, in its absence, specific DNA scars accumulate in the genome that are the resultants of the error-prone repair of double-strand DNA breaks that occur due to, for example, replication stress. There are HRD-induced DNA aberrations ranging from single nucleotide variation–based mutations to large scale, Mb sized genomic rearrangements (7). These have been combined into predictive HRD markers both in the CLIA (8) and experimental (9) setting.

These methods have proven to be a robust indicator of the absence of BRCA1/BRCA2 function both in the genomic analysis of clinical samples and in direct induction experiments (8–12). A recent study suggested that germline BRCA2 mutation-associated prostate cancer possesses some of the previously described HRD-associated mutational signatures in whole-genome sequencing (WGS) data (13).

In the current study, we investigated three related questions: (i) whether BRCA1 or BRCA2 mutations are consistently associated with the HRD-induced mutational signatures in a large set of whole-genome and whole-exome sequencing (WES) data, (ii) whether there are prostate cancer cases with HRD-associated mutational signatures that are not associated with germline or somatic BRCA1 or BRCA2 mutations, which would suggest that there might be prostate cancer cases with sensitivity to PARP inhibitor treatment even in the absence of mutations in key HR genes, and (iii) whether WGS- and WES-based methods are equally accurate to identify HRD prostate cancer cases.

In this study, we analyzed 311 whole genome sequenced samples from 240 cases (215 with localized, 13 with metastatic disease, and 12 unknown status) from the following cohorts (Table 1). For the PRAD-CA, EOPC-DE, PRAD-UK cohorts, processed data were used in our analyses. In case of the other 4 cohorts, we worked with the BAM files.

To evaluate the signs of HRD, further 498 WES samples were analyzed from the TCGA. The normal and tumor BAM files were downloaded via the GDC Data Portal.

For the PRAD-CA, EOPC-DE, PRAD-UK cohorts, germline mutations and structural variants were obtained from the DCC Data Portal. Somatic mutation status, allele-specific copy number, and structural variant data were accessed from ref. 14.

In case of the other four cohorts, germline mutations were called with HaplotypeCaller, whereas somatic point-mutations and indels were called using Mutect2 (GATK 3.8). The high fidelity of the reported variants was ensured by the application of additional hard filters on top of the tools' default ones. Allele-specific copy-number profiles was estimated by using Sequenza (15). Structural Variants were called using BRASS (v6.0.0; https://github.com/cancerit/BRASS). Further details are available in the Supplementary Notes.

The mutations were annotated using InterVar (16). Variants predicted as pathogenic or likely pathogenic were considered deleterious, whereas variants with unknown significance were marked differently. Copy-number status of BRCA1/2 was based on Sequenza results.

Somatic point-mutational signatures were determined with the deconstructSigs R package (17), by using the cosmic signatures as a mutational-process matrix. The extraction of rearrangement signatures was executed as described previously (7).

The calculation of the genomics scar scores [loss-of-heterozygosity (LOH; ref. 18), large scale transitions (LST; ref. 19), and number of telomeric allelic imbalances (ntAI; ref. 20)] were determined using the scarHRD R package (21).

Because of the lack of sufficient numbers of bona fide HRD cases within the prostate cancer cohorts, and to the dissimilarities among the cases involved in their corresponding studies, a prostate-specific HRDetect model could not be created. Instead, the weights of the original, breast cancer–specific, whole genome–based HRDetect model (9) were used to calculate the HRDetect scores of the WGS prostate samples. The scores of the WES samples were calculated by using the weights of a whole exome–specific model, that was trained on 560 artificial whole exomes. The predictors were log-transformed and standardized within all prostate cancer cases (n = 311 for WGS, and n = 498 samples for WES). Both the scores and sample attributes are available in Supplementary Tables S1 and S2.

From the 311 samples, 25 samples (from 18 patients, 3 with metastatic and 15 with localized disease) had somatic or germline BRCA1/BRCA2 mutations. Two patients (DO51087 and DO51965), carried BRCA1 germline mutations. One of the patients (DO51965, represented by five distinct sequenced biopsies) also had LOH, whereas the other (DO51087) had a somatic mutation in BRCA1. Another patient (DO52504) had a somatic structural variant aberration in BRCA1 (BRCA1-SAFB translocation).

Five patients had BRCA2 germline mutations. Of these, two had additional BRCA2 aberrations (either LOH or structural variant). Nine patients had a somatic BRCA2 mutation, out of which three also had a copy-number loss and one patient had a somatic BRCA1 mutation.

Two samples had a deep deletion in BRCA2 (Supplementary Table S1).

Both germline and somatic BRCA1/BRCA2 mutations may be present without additional loss of copy, resulting in retained homologous recombination capacity. All in all, we considered 15 biopsies from 9 cases (6 localized and 3 metastatic) as deficient either in the BRCA1 or BRCA2 function. (For summary of germline and somatic mutations see Supplementary Fig. S10).

The first scoring systems to quantify the degree of HRD were based on data derived from hybridization microarrays such as SNP arrays. The HRD-LOH score (18) is the number of LOH regions exceeding 15 Mb in size but less than the whole chromosome; the LST score (19) is defined as the number of chromosomal breaks between adjacent regions of at least 10 Mb, with a distance between them not larger than 3 Mb; and the ntAI (20) is the number of AIs (unequal contribution of parental allele sequences) that extend to the telomeric end of a chromosome. These were later adapted to next-generation sequencing (NGS) data (21). The sum of these scores is referred to as HRD score, as in previous publications (8).

In the WGS-based cohort, 9 of the 15 BRCA1/2-deficient cases had an HRD score above 42, which is the threshold of HRD previously published for ovarian cancer (8). However, there were 12 cases without any apparent BRCA1/2 or other HR gene mutation that also had a ≥42 HRD score (Fig. 1A). For the individual components of the HRD score see Supplementary Fig. S1.

NGS revealed that loss of function of BRCA1 or BRCA2 is associated with a range of distinct mutational signatures that can be extracted from NGS data. Those include: (i) a single nucleotide variation–based mutational signature (“COSMIC signature 3” or “BRCA signature” as labeled in the original publication; ref. 22); (ii) a short insertions/deletions-based mutational profile, often dominated by deletions with microhomology, a sign of alternative repair mechanisms joining double-strand breaks in the absence of homologous recombination (11); and (iii) large scale rearrangements such as non-clustered tandem duplications of a given size range (mainly associated with BRCA1 loss of function) or deletions in the range of 1–10 kb (mainly associated with BRCA2 loss of function; ref. 7). All these can be efficiently induced by the inactivation of BRCA1, BRCA2, or several other key downstream HR genes (XRCC2, XRCC3, RAD51, etc.; ref. 12).

Most BRCA-deficient prostate cancer cases showed higher levels of all three types of HRD-induced mutational aberrations (detailed results are presented in Supplementary Figs. S1–S5).

A set of consistent and easily detectable mutational features of HRD, especially in a BRCA2-deficient background, include the elevated Deletions/Insertions ratio, the relative increase in the number of deletions that are at least 10 bp long and the increased number of microhomology-mediated deletions. Such deletions are mechanistically explained as the result of alternative end-joining repairing double-strand breaks in the absence of HR (23); therefore, we considered these aberrations as the most direct evidence of HRD. Fourteen of the 15 BRCA1/2-deficient cases showed elevated levels of these features (Fig. 1B–D), consistent with previous observations (13). Notably, there were also 13 cases with likely intact BRCA1 and BRCA2 function that also had a higher deletions/insertions ratio than the BRCA1/2-deficient cases. Three of the cases had a similar fraction of microhomology-associated cases to those with BRCA2 deficiency and two of those three cases had a similar ratio of >10 bp deletions to those with BRCA2 deficiency as well. This suggests that there are BRCA1/2 intact cases that display the mutational signatures usually associated with BRCA deficiency in prostate cancer.

BRCA1 mutations are known to be associated with a large contribution of Single Nucleotide Signature 3 (24). On the basis of their signature composition, several cases (such as the DO51965_M1-M5 or DO51962_M1-M3 samples) clearly demonstrate the signs of BRCA1/2 deficiency (Supplementary Fig. S2).

Rearrangement Signature 3 and 5 have also been associated with BRCA1 and BRCA2 deficiency, respectively (7). These signs were also clearly detected in the investigated prostate cancer samples (Supplementary Fig. S3).

It was shown recently that a composite mutational signature, HRDetect, that combines the various types of mutations listed above represents a more accurate measure of HRD than any of the mutational features alone (9, 10).

The HRDetect score was determined as described in the Materials and Methods section). Fourteen out of the 15 biallelic BRCA1/2 mutants among the WGS samples had larger than 0.7 HRDetect scores (Fig. 1E), which is the currently accepted threshold of HRD samples on whole genome sequences (9, 10). One BRCA1 mutant sample (DO51087) with both a germline and a somatic mutation acquired a low, close to zero HRDetect value, which can be explained with the likely assumption that the somatic structural variant aberration also affected the germline allele, leaving the other allele intact, hence the true genotype being ±. Given that BRCA1/2 biallelic mutants are in effect HRD, the HRDetect model had almost perfectly identified the known samples with HRD. These samples, however, are not the only ones that have reached the higher ends of the HRDetect spectrum. Seventeen out of the 286 samples with a BRCA1/2 wild-type background also reached above the 0.7 threshold, further strengthening the assumption that even samples without mutations in the BRCA1/2 genes can exhibit the signs of HRD. Most of these cases had a high number of deletions with flanking microhomology, as a direct evidence of error prone correction of double-strand breaks (Supplementary Fig. S4).

Although in general the HRD score and HRDetect values showed good correlation on the entire WGS cohort (Supplementary Fig. S7), it should be noted that HRDetect was more accurate identifying the BRCA-deficient cases (Supplementary Fig. S8).

Whole-exome sequences contain approximately 1% of the sequence information contained in their whole genome–derived counterparts. In the entire set of WGS samples, the average number of single nucleotide variations per sample was 3,824. The same number for the entire WES dataset was 64 SNVs/sample. In the WGS data, the average number of deletions per sample was 256 deletions (73 with microhomologies), whereas in the WES data the same number was 4 deletions (0.75 with microhomologies)/sample (Supplementary Figs. S4–S6). Despite this significantly reduced amount of information, we attempted to derive predictors for HRD as previously described (25).

BRCA1/2 structural rearrangements, which cannot be detected by WES analysis, have been shown to underly HRD in prostate cancer (13). Therefore, we expected the number of identifiable BRCA1/2 mutants might be somewhat lower in the WES cohort. In the investigated cohort (n = 498), we found three clearly BRCA2-deficient and one BRCA1-deficient cases. Only one of those 4 cases had a ≥42 HRD score in the WES dataset (Fig. 2A). In addition to the 4 above described BRCA1- or BRCA2-deficient cases, we identified 4 cases with a deep BRCA1 deletion and 7 cases with a deep BRCA2 deletion. Three of the BRCA2 deep deletion cases also had ≥42 HRD score. Nevertheless, the majority of the likely BRCA1- or BRCA2-deficient cases or deep deleted cases had a less than 42 HRD score. This suggests that the WES-based HRD score has a significantly lower accuracy to identify HRD cases than its WGS-based counterpart.

A modified, WES-based HRDetect score was also determined for these cases (Fig. 2B).

Because a subset of the analyzed samples in the TCGA dataset had both WES and WGS data (n = 20), we performed a comparative analysis on the HRD-related attributes of the corresponding WES-WGS pairs (Supplementary Fig. S9). We found that the sum of the three genomic scars had a strong linear correlation (r = 0.79 ± 0.14) between the sets; however, this correlation was mainly driven by the number of large-scale transitions (r_LST = 0.92 ± 0.09), whereas both the HRD-LOH and TAI scores exhibited weak positive correlation (r_HRD-LOH = 0.22 ± 0.23 and r_TAI = 0.24 ± 0.23, respectively). Similar weak correlations were observed between the WES- and WGS-derived SNV signature 3 ratios (r = 0.21 ± 0.23) and microhomology-mediated deletion ratios (r = 0.23 ± 0.23; Supplementary Fig. S9). The WES-based HRDetect score was more accurate to reidentify the likely HRD cases than the WES-based HRD score because all four BRCA1- and BRCA2-deficient cases and three BRCA2 deep deleted cases had a >0.7 HRDetect score (Fig. 2B). However, there were also 78 BRCA 1/2 wild-type samples above the same HRDetect threshold that we used for the WGS measurement (>0.7). Unlike in the WGS HRDetet dataset, most of the WES-based cases with >0.7 HRDetect score had no or low number of deletions (and in most cases no microhomology flanked deletions) detected (Supplementary Fig. S6).

This lower accuracy of WES-based methods to identify HRD cases is similar to the results of a recent publication (26), in which WES was performed on BRCA-deficient and BRCA wild-type prostate cancer cases and the LST (19) and signature 3 values and their combined score as a measure of HRD were determined. In this cohort, about one third of the verified BRCA-deficient prostate cancer cases had a WES-based HRD measure in the range of the BRCA wild-type biopsies suggesting a significant rate of false negative cases.

We showed recently that in addition to eliminating BRCA1 or BRCA2, deletion of any of a number of other key HR genes (RAD51C, XRCC2, XRCC3, PALB2, RAD54) also induces the same mutational signatures usually associated with HRD (12). Therefore, we investigated whether the high HRDetect scores in BRCA 1/2 wild-type cases could be explained by the presence of biallelic aberration in other HR-related genes (such as BLM, PALB2, RAD50, RAD51 gene family, RAD52, RAD54B, and RAD54L). None of the WGS cases with higher than 0.7 HRDetect scores presented biallelic aberration in these genes. From the 55 WES cases that had higher than 0.7 HRDetect scores, 7 had biallelic BRCA1/2 mutations and an additional 3 cases could be explained by biallelic mutations in PALB2, RAD51, or RAD54B (Supplementary Figs. S10 and S11).

Ten patients had multiple tumor samples sequenced. Eight of them had HRDetect values less than 0.3 in all of their specimens. For one patient, although the HRDetect scores varied between the tumor samples, they remained below 0.7. In case of a tumor without biallelic loss of BRCA1/2 (EOPC-035), five out of the six tumor samples showed signs of HRD (Supplementary Fig. S12).

Ten patients from the investigated cohorts with metastatic prostate cancer had multiple metastatic sites sequenced. For two of the patients (A17 and A34), the primary prostate cancer and all of the metastatic sites showed signs of HRD and had higher than 0.9 HRDetect scores. All of these samples showed biallelic BRCA loss (A17: BRCA1 germline mutation and LOH, A34: BRCA2 structural variant and LOH). For six of the patients (A10, A12, A21, A22, A24, A31) the primary tumor and all of the metastatic sites had low HRDetect scores, although in some samples (A21, A22, A31) slightly elevated HRDetect scores could be observed. Two patients (A29 and A32) showed heterogeneity in HRD, where although the primary prostate cancer was HR-competent, the metastases showed signs of HRD (Supplementary Fig. S13).

Personalized therapy of prostate cancer has entered a new phase since the demonstration of the clinical efficacy of PARP inhibitors and platinum in cases with mutations in the DNA repair pathway, especially homologous recombination (3, 27). However, some of the DNA damage checkpoint gene germline mutants do not benefit from PARP inhibitor therapy (28) and some cases without such mutations are sensitive to platinum-based therapy (27, 29). Therefore, the optimal use of these therapeutic approaches will require more accurate identification of HRD cases. A similar diagnostic problem was addressed in the case of breast and ovarian cancer using HRD-induced mutational signatures (6, 8, 20). We decided to apply the same approach here. First, we demonstrated that bona fide, key HR gene (BRCA1/2)-deficient cases show all three major types of HRD-induced mutational signatures. Second, we also found that there is a significant number of cases without germline or somatic mutations in these genes that also demonstrate the same HRD-induced mutational signatures. This strongly suggests that there is an additional, identifiable set of prostate cancer cases that will likely benefit from PARP inhibitor–based therapy. In support of this assumption, a recent clinical report found at least one case without germline mutations of HR genes with exceptionally good response to platinum-based therapy showing HRD⁺ score based on WES analysis (29). A clinical trial, prioritizing patients based on such molecular signatures will be needed to further validate this hypothesis.

We could explain the presence of HRD-induced mutational signatures with mutations in other HR-related genes, such as PALB2, RAD51, or RAD54B, in a few cases. HRD can also be caused by the low expression of HR-related genes (5). However, detecting correlations between low expression of HR genes and HRD-induced mutational signatures is severely limited by factors such as normal tissue contamination. For example, significant expression deficiency or LOH of the BRCA1 or BRCA2 genes can often be masked by the presence of these genes in the normal cells in the tumor biopsy. These results are highly consistent with a recent publication that determined HRDetect values from 254 triple-negative breast cancer cases and only a subset of the high HRDetect cases could be explained by direct mutational analysis of HR-related genes (10).

Mutational signatures can be most accurately determined from WGS data because, for example, large-scale rearrangements can be detected only in such data (10). Furthermore, structural rearrangements in and around the BRCA1/2 genes, detectable only by WGS, also contribute to HRD in prostate cancer (13). Because WGS data are not always available, we investigated whether WES could also serve as a reliable detector of HRD. WES data were also able to provide a measure of HRD of clinical cases but the correlation between WGS- and WES-based HRD measures was rather limited. This is likely due to the fact that the total number of HRD-induced mutations per sample is about 100-fold lower in WES data with the number of deletions near or below the detection threshold. Therefore, one of the mechanistically strongly linked mutational signatures to HRD, microhomology-mediated deletions, cannot be reliable quantified in WES data. The limited accuracy of WES-based HRD measures was also seen in a recent publication that WES sequenced a large number of BRCA1/2 mutant and BRCA1/2 wild-type cases (26). The HRD measures of about 30% of bona fide BRCA1/2 mutant prostate cases were in the range of the BRCA1/2 wild type, likely HR-proficient cases. The higher accuracy to identify truly HRD cases by WGS data was also demonstrated recently for breast cancer (10).

Despite the wide availability of NGS, no CLIA certified WGS- or WES-based methods are available for determining HRD in the diagnostic setting. Currently, only targeted-sequencing–based estimation of the genomic scar scores are available in CLIA-certified setting such as the myChoice assay (Myriad Genetics) for breast and ovarian cancer. With this method, tumors are considered HRD⁺ if they have a high myChoice HRD score (≥42) or a tumor BRCA1/2 mutation and HRD⁻ if they have a low myChoice HRD score (<42) and wild-type BRCA1/2 (8). Unfortunately, our analysis did not have the sufficient number of BRCA-mutant cases to determine a prostate cancer–specific cutoff value, as in the case of ovarian cancer (8). Nevertheless, the BRCA-deficient cases have clearly higher genomic scar scores both in the WES and the WGS cohorts (Supplementary Figs. S5–S7). A significantly higher number of cases will be needed to determine the clinically useful threshold value. Using the ovarian cancer–specific cutoff value (≥42), the WGS-based HRD score was reasonably accurate to identify HRD cases but it was outperformed by WGS-based HRDetect.

It has been previously demonstrated that (30) metastasis to metastasis spread of prostate tumor clones is common. This may explain the small heterogeneity seen in HR status of multiple metastases.

In summary, we show that the presence of BRCA1/2 mutations is associated with the presence of HRD-associated mutational signatures in prostate adenocarcinoma. HRD patterns were also detected in a subset of patients who did not harbor germline or somatic mutations in BRCA1/2 or other HR-related genes. This likely defines a subset of patients with prostate cancer, most accurately identified by WGS data, that will likely benefit from PARP inhibitor or platinum-based therapy.

Z. Szallasi is listed as a co-inventor on a patent to quantify homologous recombination deficiency, which is owned by Children's Hospital Boston and licensed to Myriad Genetics. No potential conflicts of interest were disclosed by the other authors.

The results shown here are based upon data generated by the TCGA Research Network: http://cancergenome.nih.gov/ and the International Cancer Genome Consortium (ICGC): https://icgc.org/

Conception and design: Z. Sztupinszki, M. Diossy, M. Krzystanek, M.M. Pomerantz, V. Tisza, S. Spisak, I. Csabai, Z. Szallasi

Development of methodology: Z. Sztupinszki, M. Diossy, Z. Szallasi

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Z. Sztupinszki, M. Krzystanek, M.M. Pomerantz

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Z. Sztupinszki, M. Diossy, J. Borcsok, M.L. Freedman, Z. Szallasi

Writing, review, and/or revision of the manuscript: Z. Sztupinszki, M. Diossy, M. Krzystanek, J. Borcsok, M.M. Pomerantz, O. Rusz, I. Csabai, M.L. Freedman, Z. Szallasi

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Z. Sztupinszki, Z. Szallasi

Study supervision: Z. Szallasi

This work was supported by the Research and Technology Innovation Fund (KTIA_NAP_13-2014-0021 and NAP2-2017-1.2.1-NKP-0002); Breast Cancer Research Foundation (BCRF-17-156 to Z. Szallasi) and the Novo Nordisk Foundation Interdisciplinary Synergy Program Grant (NNF15OC0016584), Det Fri Forskningsrad (award number #7016-00345B; to Z. Szallasi); Department of Defense through the Prostate Cancer Research Program (award number is W81XWH-18-2-0056; to Z. Szallasi and M.L. Freedman). The Danish Cancer Society grant (R90-A6213; to M.L. Freedman). Z. Szallasi and J. Borcsok were supported by Velux Foundation 00018310 grant.

Data Acknowledgement Statement: This study was based on the analysis of dbGaP data sets phs000661.v1.p1, supported by grant #HG003079 through NHGRI. and phs001105.v1.p1., supported by Intramural Research Program of the National Human Genome Research Institute and Prostate Cancer SPORE Grant # CA17205. We thank the Mayo Clinic Pathology Core for sample processing. This work utilized the National Institutes of Health Biowulf high-performance computing cluster. Finally, we thank the many men who provided samples for this study.

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.