BRCA2, ATM, and CDK12 Defects Differentially Shape Prostate Tumor Driver Genomics and Clinical Aggression

Defects in cellular pathways responsible for DNA damage repair (DDR) are common across cancer types and can accelerate tumor progression due to the continual accrual of secondary genomic alterations (1). In metastatic prostate cancer, germline and/or somatic DDR defects can be identified in 15%–30% of patients (2, 3). However, it is increasingly recognized that individual DDR defects are not equal in their genomic or clinical consequences (4–6). Deeper characterization is required for accurate prognostication of patients and optimal treatment selection.

The most commonly altered DDR genes in metastatic prostate cancer are BRCA2, ATM, and CDK12 (7). Deleterious BRCA2 alterations result in failure of homologous recombination repair (HRR) and increased tumor cell reliance on compensatory pathways that can be more ‘error-prone' when repairing double-strand DNA breaks (8). Consequently, BRCA2-defective genomes bear distinctive mutational signatures (4, 6, 9). Defects in CDK12 do not associate with signatures of HRR failure and are instead accompanied by a genome-wide pattern of focal tandem duplications (5, 6, 10). ATM loss has not yet been linked with a genomic phenotype and does not appear to have a direct role in HRR (11). Surprisingly, while BRCA2, ATM, and CDK12 defects can have drastically different genome-wide consequences, any association with downstream prostate cancer driver gene alterations has not been explored.

Some DDR defects are associated with aggressive prostate cancer (12). For example, BRCA2 and CDK12 mutations and/or deletions are rare in localized compared with metastatic prostate cancer (13), and both genes have been linked to short overall survival (OS) in the castration-resistant setting (10, 12, 14, 15). However, there is debate over the prognosis and outcomes of patients with DDR defective metastatic prostate cancer with standard-of-care androgen receptor (AR)-targeted and chemotherapies (16). Regardless, HRR defects can be exploited by therapies that induce more DNA damage and/or inhibit the compensatory DDR pathways (8). Preclinical studies and clinical trials in several cancer types have demonstrated that cells defective in BRCA1 or BRCA2 are highly sensitive to inhibition of poly (adenosine diphosphate-ribose) polymerase (PARP; ref. 17). In prostate cancer clinical trials to date, treatment efficacy in ATM- and CDK12-defective cancers appears weaker than for BRCA2 (7, 18).

In this study, our objectives were to: (i) estimate the frequency of biallelic BRCA2, ATM, and CDK12 defects in liquid biopsies from men with metastatic prostate cancer, (ii) define the somatic driver gene alterations associated with each DDR defect, and (iii) explore the clinical characteristics and prognoses of BRCA2-, ATM-, and CDK12-defective metastatic prostate cancers.

We analyzed targeted sequencing data from 1,615 plasma cell-free DNA (cfDNA) and matched leukocyte DNA samples from 879 consecutive patients with metastatic prostate cancer enrolled in a liquid biopsy program at the Vancouver Prostate Centre and BC Cancer; final data freeze occurred in August 2019. Patients were unselected for age at cancer diagnosis or familial cancer history. Approval for collection and profiling of patient samples was granted by the University of British Columbia Research Ethics Board (certificate numbers H18-00944, H16-00934, and H09-01628). The study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all patients prior to enrollment.

Initial targeted sequencing of metacohort cfDNA and leukocyte DNA was conducted using a previously published targeted panel capturing the exons of 73 prostate cancer relevant genes, including 22 genes associated with DDR: ATM, ATR, BRCA1, BRCA2, CDK12, ERCC1, ERCC2, ERCC3, ERCC4, ERCC5, FANCA, FANCC, FAND2, FANCE, FANCF, FANCG, MLH1, MSH2, MSH6, PALB2, RAD51B, and RAD51C (15, 19–21). We retrospectively queried the metacohort for patients harboring germline and/or somatic tumor variants in at least one of these genes. Patients were included in the final study cohort if they had sample(s) with one or more deleterious somatic or germline defects in BRCA2, ATM, and/or CDK12 (Supplementary Materials and Methods). Only cfDNA samples with ≥2% tumor-derived cfDNA fraction were used for somatic mutation analyses. Patients with somatic homozygous deletions in these three genes were also included in the study cohort. Note that copy-number changes were called in the subset of samples with ≥20% tumor-derived cfDNA fraction. The majority of study cohort patients had ≥1 cfDNA sample collected after progression to castration resistance, although 5 patients only provided samples in the castrate-sensitive setting (Supplementary Table S1).

Plasma cfDNA and germline DNA library construction and targeted panel hybridization was performed according to previously validated methodology (15, 19–21). A subset of libraries were sequenced on a second custom panel covering the full gene lengths (including flanking sequences and intronic regions) of key tumor suppressors (Supplementary Table S2). All purified library pools were sequenced on an Illumina MiSeq or HiSeq 1500/2500 machine. Identification of somatic variants, rare germline variants, and copy-number alterations, as well as estimation of circulating tumor DNA (ctDNA) purity, was performed using published methods (15, 19–21). Categorization of mutant allele–specific imbalance [heterozygous loss, loss-of-heterozygosity (LOH), or nondeletion LOH] in BRCA2, ATM, and CDK12 was accomplished via review of gene coverage log ratio and heterozygous single nucleotide polymorphism (SNP) imbalance (Supplementary Materials and Methods).

Clinical characteristics and outcomes in the study cohort with DDR defects were collected retrospectively (Table 1). Time to castration resistance was calculated from androgen deprivation therapy (ADT) initiation (or date of bilateral orchiectomy in 1 patient P102). OS was defined as start of first-line metastatic castration-resistant prostate cancer (mCRPC) therapy to death or last follow-up. Outcomes and genomic features were compared with a control cohort of 187 patients with mCRPC from a randomized phase II study (Supplementary Materials and Methods; refs. 15, 22). Clinical endpoint definitions adhered to Prostate Cancer Working Group 3 criteria and were consistent between study and control cohorts (23). Survival functions were estimated via the Kaplan–Meier method and compared using the log-rank test. Descriptive statistics constituted the majority of statistical analyses: Fisher's exact test and the Mann–Whitney U test were used for cross-categorical comparisons. All hypothesis tests were two tailed and required a 5% significance threshold.

Identifying somatic DDR defects is challenging in patients with metastatic prostate cancer. Fresh metastatic tissue is impractical to obtain in routine clinical practice, and failure rates approach 40% for analysis of archival diagnostic tissue. An alternative approach is to leverage ctDNA, which is abundant in 60%–80% of patients with metastatic disease (15, 19, 24). Here, we performed a retrospective analysis of plasma cfDNA and matched leukocyte DNA-targeted sequencing data from 879 men with metastatic prostate cancer—including 800 (91%) with castration-resistant disease—sequenced to median unique read depths of 964× (cfDNA) and 500× (leukocyte DNA; Fig. 1A). A total of 57 patients [6.5%; 95% binomial proportion confidence interval (CI): 4.9–8.3] harbored a deleterious germline frameshift (or in one case, missense) mutation in one or more DDR genes (Fig. 1B). Deleterious BRCA2 and ATM germline mutations were the most prevalent, present in 34 (3.9%) and 6 (0.7%) patients, respectively, but eight truncating mutations were identified in members of the Fanconi anemia complex (FANC). CHEK2 was not included in our targeted capture panel and could not be assessed. Germline mutation frequencies for ATM and BRCA2 were comparable with previously published cohorts (Supplementary Table S3).

Identification of both somatic and germline mutations was performed in the 635 of 879 patients where ctDNA fraction (i.e., tumor content) was greater than 2% in at least one cfDNA collection (Fig. 1A). In total, 20.5% (95% binomial CI: 17.4–23.8; 130 patients) harbored a deleterious germline and/or somatic mutation in at least one DDR gene. Of these 130 patients, 21 (3.3% of 635) exhibited somatic mismatch repair defective (MMRd) etiology and were excluded from subsequent analysis because these tumors are predisposed to accruing temporally variable passenger and/or subclonal mutations (including in DDR genes; ref. 21). Therefore, 109 patients (17.2%; 95% binomial CI: 14.3–2.3) had at least one deleterious somatic and/or germline mutation in a DDR gene (excluding MMRd; Fig. 1C; Supplementary Fig. S1). BRCA2 (7.3%), ATM (3.0%), and CDK12 (4.4%; 93/109 patients total) were the most commonly affected genes (Fig. 1C) and no other DDR gene on our targeted panel was mutated in more than 0.6% of patients (4/635). The majority (96%) of deleterious somatic alterations in BRCA2, ATM, and CDK12 were estimated to be clonal; this is in contrast to the 10 potentially deleterious BRCA2, ATM, and CDK12 variants detected across the 21 MMRd tumors, of which only 60% (6/10) were putatively clonal (76/79 vs. 6/10, P = 0.0025, Fisher's exact test; Supplementary Materials and Methods). On average, patients with BRCA2-, ATM-, and CDK12-mutant tumors had higher ctDNA fractions than ctDNA-positive patients without a deleterious DDR mutation (Fig. 1D; Supplementary Materials and Methods; ref. 15).

In mCRPC tissue sequencing studies, BRCA2 homozygous deletions are identified in 2%–4% of cases (3). Because homozygous deletions are inferred from coverage log-ratio shifts, they are particularly challenging to detect in scenarios of low tumor content (i.e., low ctDNA fraction). Nevertheless, in 345 patients where ctDNA fraction was above 20% (Supplementary Materials and Methods), we detected homozygous BRCA2 deletions in 8 patients (2.3%; 95% binomial CI: 1.0–4.5; Fig. 1A; Supplementary Fig. S2). No other DDR gene demonstrated recurrent homozygous deletions. Among the 345 (non-MMRd) patients that were evaluable for both BRCA2 mutations and homozygous deletions, 10.4% (n = 36; 95% binomial CI: 7.4–14.2) exhibited deleterious alterations. This is comparable with the PROfound phase III clinical trial of PARP inhibitor olaparib where 8.7% of tumor tissue–screened patients with mCRPC exhibited deleterious BRCA2 alterations (17).

We combined the 8 patients with homozygous BRCA2 deletions together with the 99 patients harboring deleterious germline and/or somatic mutations in BRCA2, ATM, or CDK12 to form the study cohort explored henceforth (238 cfDNA samples from 107 patients, including 6 germline patients without a ctDNA-positive sample; Fig. 1E–H; Supplementary Table S1).

In prostate and other cancers, monoallelic DDR gene defects (such as single copy deletion) are infrequently associated with downstream genomic signatures of defective DNA repair (25). Therefore, we evaluated the proportion of the study cohort that had evidence of biallelic alterations, via either a second somatic mutation or LOH across the first mutation (Fig. 2A; Supplementary Tables S4–S7). Patients with homozygous deletions, and those without evidence for ctDNA >2% were excluded, leaving a total of 93 patients for evaluation. For patients that provided multiple cfDNA samples, we used the sample with the highest ctDNA fraction (Supplementary Table S1; Supplementary Materials and Methods).

A second deleterious mutation was detected in only 5 of 65 patients with a truncating mutation in ATM or BRCA2 (7.7%). Conversely, half of CDK12 patients (14/28; 50.0%) harbored a second mutation, representing a significantly higher proportion of patients than in either BRCA2 or ATM cases (P = 1.2 × 10^–5, Fisher–Freeman–Halton test; Fig. 2B–E; Supplementary Tables S4–S7). LOH across a mutant allele can occur via deletion of the non-mutant allele, but can also manifest in a copy neutral or even mutant allele–specific copy gain scenario (i.e., non-deletion LOH; ref. 26). Therefore, to assess LOH rates for each gene, we evaluated both gene copy number and the somatic deviation in allele frequency (relative to 50%) for intragenic germline heterozygous SNPs that were captured by targeted sequencing (see Supplementary Materials and Methods). Focusing on samples with sufficient ctDNA to detect perturbations in allelic configuration and without a second somatic mutation, 62.1% (18/29) of BRCA2 and 41.7% (5/12) of ATM cases showed evidence for LOH via monoallelic deletion (Fig. 2B–E; Supplementary Fig. S1; Supplementary Tables S4–S7). Conversely, LOH by monoallelic deletion was rare in the context of a CDK12 mutation (one case only; 6.3%), although copy neutral LOH was observed in two CDK12 cases (Fig. 2B–E; Supplementary Tables S4–S7). A third CDK12 case showed mutant allele–specific imbalance caused by copy gain of the mutated allele. Overall, these data show that biallelic inactivation of DDR genes is common in metastatic prostate cancer, but that the mechanisms of disruption differ by gene.

To test whether metastatic prostate cancers with DDR defects are associated with specific classes of additional driver alterations, we compared the frequency of driver gene alterations in the study cohort (Fig. 3A) with ctDNA-positive mCRPC patients without deleterious BRCA2, ATM, or CDK12 defects from a recent randomized phase II clinical trial (Supplementary Tables S4–S6; Supplementary Materials and Methods). To reduce false negatives in our copy-number variant (CNV) frequency comparisons, we restricted analysis to samples with ctDNA fractions above 20% and 5% for autosomal CNVs and CNVs affecting the AR, respectively (Supplementary Materials and Methods).

CDK12-mutant samples displayed recurrent copy-number gain of oncogenes (Fig. 3A and B; Supplementary Fig. S3), matching the focal tandem duplication pattern observed in prostate and ovarian cancer from tissue whole-exome sequencing (5, 27). Copy gains of cell-cycle driver CCND1 were more common in patients with CDK12 mutations than in all other categories (Fig. 3B; P = 4.3 × 10^–6; Fisher's exact test), as reported previously (5). The TP53-specific ubiquitin ligase MDM2 showed copy gain in 6 of 16 CDK12 cases, but only 1 of 64 DDR-intact cases [OR = 37.8 (95% CI: 4.1–347.9); P = 1.7 × 10^–4; Fisher's exact test]. Conversely, alterations in tumor suppressors RB1, TP53, and PTEN were significantly less common in CDK12-defective tumors than in DDR-intact tumors (Fig. 3C). Copy gains of AR were detected in all but three CDK12 cases (22/25; Fig. 3D), although the median number of AR copies in CDK12 cases with AR gain (after adjustment for ctDNA fraction) was lower than in DDR-intact controls (4.8 vs. 8.7; P = 9.1 × 10^–3; Mann–Whitney U test; Supplementary Fig. S4). Interestingly, lower absolute AR copy number was also observed in BRCA2-defective tumors (3.7 vs. 8.7; P = 2.3 × 10^–4; Mann–Whitney U test), and this result remained significant after restricting comparisons with patients with ctDNA fraction above 20% (3.5 vs. 8.9; P = 7.3 × 10^–4; Supplementary Fig. S4). Lower AR copy number was not identified in cases with ATM truncating mutations. We recapitulated these associations using the study cohort subset with evidence for biallelic disruption (Supplementary Fig. S5).

Collectively, samples with BRCA2, ATM, and CDK12 defects exhibited significantly fewer point mutations and indels in TP53 than DDR-intact controls [23.7% vs. 47.2%; OR = 0.4 (95% CI: 0.2–0.6); P = 3.1 × 10^–4; Fisher's exact test]. No significant differences between groups were observed for TP53 gene body deletions, although the number of evaluable CDK12- and ATM-mutant samples was small (Fig. 3E; Supplementary Fig. S5). For CDK12-mutant samples, the low TP53 mutation frequency may be explained by the oncogene tandem duplication phenotype (including MDM2 copy gain; Fig. 3B). Conversely, ATM is a kinase that mediates phosphorylation of TP53 in response to DNA damage, and mutations in these two genes appear to be epistatic in breast cancer (28). Here, and in public metastatic prostate cancer data, ATM and TP53 mutations also appeared mutually exclusive (ref. 3; Fig. 3F and G).

Because BRCA2 mutations result in higher rates of structural variation than in typical prostate cancer (4, 5), we hypothesized that an increase in disruption to TP53 via large structural rearrangements may account for the low rate of observed TP53 mutation. Therefore, in 44 patient samples with BRCA2 defects (median ctDNA fraction 35.5%), we performed targeted sequencing across the introns and flanking regions of TP53, as well as several other tumor suppressor genes (Supplementary Tables S2 and S8; Supplementary Fig. S6). We identified 10 structural rearrangements affecting TP53 in 9 BRCA2 patients; none were previously identified via log-ratio copy-number analysis (Fig. 3H; Supplementary Table S8). BRCA2 is located on chromosome 13, 5 Mbp from RB1. The two genes are reported to be frequently co-deleted, and dual disruption is linked to aggressive biology (29). Consistent with this finding, over 70% of our evaluable BRCA2-defective cases exhibited RB1 deletion, significantly more than DDR-intact controls [70.3% vs. 45.3%; OR = 2.9 (95% CI: 1.2–6.7); P = 2.2 × 10^–2; Fisher's exact test; Fig. 3C]. Structural rearrangements were also very common across the RB1 gene locus, identified in 14 of 44 (32%) BRCA2 patients with intron sequencing data, and predominantly manifesting as large intragenic deletions (median size: 1.18 kb) with intronic breakpoints (none of which were identified via prior analysis of the average log-ratio signal across all RB1 exons). Three patients (P86, P46, and P23) harbored truncating interchromosomal translocations involving RB1, and a fourth (P50) exhibited a 10 kb intragenic tandem duplication disrupting RB1 exons 19–23, consistent with a previous report of mechanistic diversity underlying RB1 genomic inactivation in mCRPC (ref. 30; Supplementary Table S8; Supplementary Fig. S6). Together, these data point toward frequent loss of tumor suppressors via unbalanced structural rearrangements (rather than mutation) in cases with BRCA2 defects, compatible with the characteristic HRR-deficient genotypic signature.

To test whether the detection of DDR gene mutations is consistent over time in ctDNA, we analyzed the 61 patients in the study cohort who provided serial cfDNA samples during the course of their treatment for metastatic prostate cancer (192 samples total; median 3 per patient; Supplementary Table S1). Samples represented an average disease course of 10.5 months (interquartile range: 9.7 months, median 2.3 months/sample), with the majority collected in the castration-resistant setting. In this subset of patients, we had identified a total of 79 deleterious somatic mutations in BRCA2 (n = 18), ATM (n = 19), and CDK12 (n = 42), via our prior analysis of the cfDNA sample with the highest ctDNA fraction (Fig. 2C; Supplementary Tables S4–S7). The average independent DDR mutation re-detection rate across all other cfDNA samples with evidence of ctDNA was 83% (Fig. 4A). Remarkably, this rose to 94% when leveraging a priori information to recover variants with mutant-read support below our conservative independent detection threshold (Fig. 4A; Supplementary Materials and Methods).

We next asked whether somatic DDR gene alterations identified via ctDNA analysis in the metastatic setting were present in patient-matched tumor tissue. A total of 22 of 107 patients had an evaluable archival tissue sample available for analysis, predominantly reflecting systemic therapy-naïve primary disease (Supplementary Table S9; Supplementary Materials and Methods). There were 15 somatic mutations in ATM, BRCA2, or CDK12 identified in ctDNA from these patients (some patients with tissue had only germline mutations and/or copy-number alterations). Using the identical targeted sequencing approach as applied to cfDNA, 100% (15/15) of somatic mutations were independently detected in primary tissue (Supplementary Table S10). There were six BRCA2-mutant cases for which biallelic inactivation could not be confirmed in their liquid biopsy (due to low ctDNA fractions); interestingly, all six cases exhibited clear copy-loss LOH in archival tissue. Collectively, these data support prior suggestions that DDR-gene inactivation is an early event in aggressive prostate cancer (31).

Reversion mutations that restore BRCA2 function are a resistance mechanism in BRCA2-mutated prostate cancers treated with PARP inhibitors or platinum-based chemotherapy (32–38). Our study cohort included 4 patients with BRCA2-deficient mCRPC exposed to one of these therapies, and who had provided plasma cfDNA samples before and after treatment. In all four cases, the post-treatment cfDNA samples exhibited insertions or deletions that restored an open BRCA2 reading frame, thereby plausibly re-establishing HRR function (Supplementary Tables S4 and S8). We synthesized our findings with previously reported cases for a total of 119 unique reversion alterations across 14 patients (median 7.5 reported per patient; refs. 32–37; Fig. 4B–E). The median size of detected secondary frameshift-correcting BRCA2 alterations was 19 bp (range: 1–2,269 bp; Fig. 4B) and consisted mostly of deletions. Most published sequencing approaches (including our own prior work) did not analyze BRCA2 introns nor perform structural variant analysis, and therefore may have under-identified large and/or complex reversions. Among our three post-treatment samples with sequencing data across BRCA2 exons and introns, we identified a total of 15 variants larger than 19 bp (median reversion size; Supplementary Tables S4 and S8). All 15 variants restored an open reading frame [probability of occurring by chance = (1/3)¹⁵ or 7.0 × 10^–8], suggesting that detection of large insertion/deletion events is synonymous with a high likelihood of them restoring BRCA2 function. Forty-one percent (49/119) of reversion alterations spanned the original deleterious somatic or germline BRCA2 variant (thereby removing it), and the majority of reversions were highly co-localized to the position of initial alteration (Fig. 4D). In several cases, large portions of the BRCA2 repeats domain were excised (Fig. 4E).

Sixty-two of 107 study cohort patients had clinical data available for retrospective exploratory outcomes analyses (Fig. 1H). Patient characteristics are summarized in Table 1 and were on average consistent with aggressive disease. Patients with DDR-defective metastatic prostate cancer had a high rate of de novo metastatic presentation, and a significantly shorter time to castration-resistant disease progression from ADT commencement versus a comparator cohort of 187 patients with metastatic prostate cancer without detected deleterious germline and/or somatic DDR defects (Table 1; Fig. 5A; ref. 22). However, these associations appeared strongest in the context of BRCA2 defects (Table 1; Fig. 5A–C). Patients with germline and/or somatic BRCA2 defects had a significantly shorter response to AR-pathway inhibitors (ARPI) for first-line mCRPC treatment, as well as significantly shorter OS versus the comparator cohort subset with >2% ctDNA fraction (Fig. 5B and C). Conversely, deleterious ATM defects were not associated with worse outcomes. CDK12 patients displayed an inferior time to PSA progression on first-line ARPIs but no significant difference in OS. Recognizing that BRCA2-defective mCRPC outcomes may be significantly altered by PARP inhibitors or platinum-based chemotherapy, we performed a subset analysis of the 12 patients in our cohort that received at least one of these treatments (2 patients received both). This subset of patients had a longer OS compared with patients with BRCA2-defective mCRPC not exposed to these agents [median 29.2 vs. 14.7 months from start of first-line mCRPC therapy; HR 0.17 (95% CI: 0.06–0.52); P = 1.7 × 10^–3; Fig. 5D].

The activity of PARP inhibitors in mCRPC with HRR defects underscores the urgent need for clinically practical tumor genotyping. Indeed, clinical guidelines for the management of metastatic prostate cancer now recommend germline and/or somatic testing for DNA repair defects (39–41). Our results from a large cohort of metastatic prostate cancer demonstrate that deleterious alterations in BRCA2, ATM, and CDK12 are detectable in plasma ctDNA at frequencies similar to those observed in the population with successful metastatic tissue biopsy (3, 4). These data support the development of minimally invasive liquid biopsy approaches to identify mCRPC cases most likely to respond to PARP inhibitors and other treatments that exploit defective DNA repair.

DDR gene status was concordant between archival primary tissue taken at cancer diagnosis and serial ctDNA-positive samples collected in the mCRPC setting. This is consistent with a recent study of patient-matched archival tissue and late-stage metastatic biopsy, where all nine DDR mutations identified in castration-resistant metastatic lesions were also present in the diagnostic tissue (31). Together, these data suggest that most DDR alterations are truncal to the metastatic lineage, and that archival tissue is likely to be a reliable alternative to fresh tissue biopsy for DDR genotyping and screening for PARP inhibitor eligibility. This is advantageous for screening efforts because prostate cancer metastatic biopsies will not be widely adopted outside of specialized academic centers. Nonetheless, up to 45% of patients may be failed by a strategy that relies solely on archival tissue, due to the challenge of external block retrieval and the often poor quality of DNA extracted from formalin-fixed paraffin embedded tissue (17). Our data, highlighting the depth of information that can be obtained from ctDNA, means we can envisage a scenario where one testing modality can offer an alternative should the other fail or provide an ambiguous result. Reliance on ctDNA alone will fail to inform on somatic alterations in the significant minority of patients with very low ctDNA fractions.

MMRd etiology underlies 3%–5% of metastatic prostate cancer and enables frequent accrual of passenger and/or subclonal mutations in a range of cancer-related genes including BRCA2 (21). Subclonal variants and those not under selective pressure are more likely to be temporally variable, and their detection may not necessarily indicate bulk oncogenic reliance on affected pathways (e.g., HRR in context of BRCA2 variants). Consequently, MMRd should be considered separately from all other DDR defects. Immune checkpoint inhibitors can be effective in this molecular subtype (2, 42).

Biallelic gene alterations were detected in ctDNA from the majority of evaluable patients with a deleterious mutation in BRCA2, ATM, or CDK12. These results mirror findings from metastatic tissue biopsies and are in line with a pan-cancer study where selective pressure for biallelic BRCA2 gene inactivation was only observed in cancers that showed frequent germline disruption of HRR genes (biallelic rate = 70% in prostate; refs. 3, 25). The observed high frequency of biallelic loss suggests that detection of a deleterious DDR mutation (but not a single copy deletion) may serve as plausible grounds to assume loss of both alleles; this working assumption would simplify DDR testing, particularly because mechanisms of biallelic inactivation differ by gene and some cannot be assessed in the context of low tumor content (i.e., low ctDNA fraction). Interestingly, we observed several cases of non-deletion LOH across CDK12 mutations. CDK12 homozygous deletions were not identified here and are extremely rare in the literature (10). One possible explanation is that MED1 (a member of the essential mediator complex) lies immediately upstream of CDK12 and may exert negative selection pressure against homozygous deletion of this locus (43, 44).

We show that deleterious defects in BRCA2, ATM, and CDK12 differentially associate with specific somatic alterations across established prostate cancer driver genes. BRCA2 tumors displayed frequent inactivation of tumor suppressors via structural rearrangement. RB1 was particularly affected, consistent with a recent study that linked co-deletion of BRCA2 and RB1 to tumor aggression (29). Indeed, loss of RB1 in context of BRCA2 may indicate a predilection for lineage plasticity and progression to neuroendocrine prostate cancer. Conversely, CDK12 defects were associated with widespread low-level copy gains of oncogenes, including the AR and MDM2 (a negative regulator of TP53). These specific observations in BRCA2- and CDK12-defective tumors are broadly compatible with their characteristic phenotypes (homologous recombination deficiency and tandem duplications, respectively; refs. 5, 9, and 27). Tumor suppressor loss is linked to poor prognosis in metastatic prostate cancer (15, 45). Therefore, frequent control of PTEN/RB1/TP53 in BRCA2- and CDK12-defective mCRPC (albeit via very different mechanisms of both direct disruption and/or modification to regulatory mediators) may partially explain the poor outcomes of these patients in our study.

Our results suggest that ATM defects associate with less clinical aggression compared with BRCA2- or CDK12-defective tumors. We observed a depletion of TP53 alterations in ATM-mutated disease relative to mCRPC without evidence of deleterious DDR defects. However, ATM also has a role in phosphorylating TP53, and thus ATM defects may act as a partial phenocopy for TP53 loss. A recent pan-cancer study showed that both monoallelic and biallelic ATM loss differentially associated with improved clinical benefit following radiotherapy, but only in the context of wild-type TP53 (46). In future, understanding the degree of functional redundancy between these two genes (and across allelic states) may help refine genomic-based prognostication in mCRPC, and better define the role ATM plays in prostate cancer DDR.

Patients in our study with BRCA2- or CDK12-defective metastatic prostate cancer had poor outcomes with historical standard of care relative to patients without evidence of DDR defects. In general, germline BRCA2 mutations are associated with poor prognosis across clinical states (12, 47, 48). Somatic CDK12 mutations are also linked consistently with poor prognosis (10, 14, 49, 50). There is less clarity for somatic BRCA2 defects, and conflicting studies are difficult to compare due to differences in patient populations, study methodology, source of tumor material, and definitions of pathogenicity (3, 15, 51, 52). In our study, patients with somatic or germline BRCA2 defects experienced shorter time to PSA progression on ARPIs for first-line mCRPC compared with ctDNA-positive patients without evidence for deleterious DDR defects (median 6.1 vs. 8.3 months). These results are compatible with data across subgroup analyses (outcomes stratified by HRR status) from a prospective phase II trial comparing abiraterone with abiraterone plus olaparib in unselected ARPI-naïve mCRPC (ref. 53; median 6.5 vs. 9.7 months radiographic progression-free survival on abiraterone monotherapy). Regardless, in the coming years, we should expect the prognosis of BRCA2-defective prostate cancer to improve, because response rates to PARP inhibitors are high in this population. Future examination of OS should stratify BRCA2 patients by PARP inhibitor and platinum-chemotherapy exposure. Patients in our study with deleterious germline or somatic BRCA2 defects treated with these agents had significantly improved OS compared with the same genotypic group treated with standard of care. We caution that selection bias is likely contributing to the magnitude of difference between these groups, as these patients were deemed fit enough to receive PARP inhibitors in a clinical trial or platinum chemotherapy.

Reversion mutations appear to be the predominant mechanism of acquired resistance to PARP inhibitors or platinum-based chemotherapy in BRCA2-defective mCRPC (32–38). Similar events have been observed in BRCA1-defective ovarian cancer (54), and both treatments appear to associate with the same reversion characteristics. The spectrum of reversion alterations in the 14 patients with BRCA2-defective prostate cancer analyzed as part of our study closely align with a recent pan-BRCA-lineage cancer analysis, supporting a lack of positional polarity relative to the pathogenic mutation, and a restricted range of reversion sizes (<100 bp in the majority cases; ref. 38). Notably, across the literature to date no reversions are yet reported in CDK12 or ATM among patients with these alterations treated with DNA-damaging drugs, supporting clinical observations that the original deleterious mutations in these genes are not as sensitizing as per BRCA2 (18). cfDNA is a favorable analyte to monitor patients receiving PARP inhibitors or platinum chemotherapy for reversion alterations, owing to its clinical practicality and ability to capture greater polyclonal diversity than biopsy of a single disease focus (32). Our results encourage the incorporation of BRCA2 introns into targeted sequencing panel design, and the use of split-read and/or discordant paired end read mapping techniques to identify large reversion mutations which could otherwise be missed.

The detection of copy-number changes in cfDNA samples with low ctDNA fraction presents a considerable challenge for assay design and bioinformatic methodology. A key limitation of our study is that we did not perform copy-number profiling in samples with <20% ctDNA due to the potential for false negatives. Orthogonal assays, such as low-pass whole-genome sequencing can help increase sensitivity for broad allelic imbalance and/or chromosomal arm level alterations but will not improve detection of focal biallelic deletions (e.g., in BRCA2). Future targeted sequencing assays covering additional genomic territory around actionable genes will have greater sensitivity for detection of biallelic deletions, but are unlikely to achieve limits of detection observed for point mutations. Subclonal copy-number alterations, which are even more difficult to detect and characterize than clonal alterations, will likely remain cryptic in all except very high ctDNA fraction samples.

This study did not address defects across rarely altered DDR genes, collectively comprising another approximately 10% of metastatic prostate cancers from recent tissue-based whole-exome and whole-genome sequencing studies (3, 4). Understanding the genomic and clinical relevance of each of these rare gene alterations in isolation (e.g., PALB2, ATR, CHEK2) will require considerably larger patient cohorts. Future screening to identify candidates for PARP inhibitors or platinum-based therapy may incorporate genome-wide mutational signatures, as this could provide evidence of compromised HRR capacity in circumstances where a causal defect cannot otherwise be identified (e.g., intronic rearrangements in exon-limited assays or epigenetic silencing). However, integration of genomic signature analysis into clinical practice will require whole-exome or whole-genome sequencing which may be prohibitively expensive for large population screening efforts.

S.M. Yip reports grants and other from AstraZeneca/Merck and other from Pfizer outside the submitted work. K.N. Chi reports grants from Janssen, Astellas, and Sanofi during the conduct of the study. Dr. Chi also reports grants and personal fees from Janssen, Astellas, Astra Zeneca, and Sanofi, as well as personal fees from Constellation Pharmaceuticals, Daiichi Sankyo, Merck, Novartis, Pfizer, Point Biopharma, and Roche outside the submitted work. A.W. Wyatt reports grants from Canadian Institutes of Health Research (CIHR), Prostate Cancer Canada, and Prostate Cancer Foundation during the conduct of the study. Dr. Wyatt also reports personal fees from AstraZeneca, Astellas, and Merck, as well as grants and personal fees from Janssen outside the submitted work. No disclosures were reported by the other authors.

E. Warner: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. C. Herberts: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. S. Fu: Resources, data curation, investigation, writing-review and editing. S. Yip: Resources, data curation, investigation, writing-review and editing. A. Wong: Data curation, formal analysis, investigation, writing-review and editing. G. Wang: Data curation, validation, investigation, methodology, writing-review and editing. E. Ritch: Data curation, software, methodology, writing-review and editing. A.J. Murtha: Data curation, software, writing-review and editing. G. Vandekerkhove: Data curation, writing-review and editing. N.M. Fonseca: Data curation, writing-review and editing. A. Angeles: Data curation, writing-review and editing. A. Beigi: Data curation, writing-review and editing. E. Schönlau: Data curation, validation, investigation, methodology, writing-review and editing. K. Beja: Data curation, validation, investigation, methodology, writing-review and editing. M. Annala: Data curation, software, formal analysis, project administration, writing-review and editing. D. Khalaf: Data curation, writing-review and editing. K.N. Chi: Conceptualization, resources, supervision, funding acquisition, project administration, writing-review and editing. A.W. Wyatt: Conceptualization, resources, supervision, funding acquisition, writing-original draft, project administration, writing-review and editing.

This work was funded by the Canadian Institutes of Health Research via a Project Grant to A.W. Wyatt and K.N. Chi, Prostate Cancer Canada via a Movember Rising Star in Prostate Cancer Research Award to A.W. Wyatt, and the Prostate Cancer Foundation via a PCF Challenge Award to A.W. Wyatt. Additional support was provided by the BC Cancer Foundation (K.N. Chi) and the Jane and Aatos Erkko Foundation (M. Annala).

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.