Genomic Analysis of Circulating Tumor DNA in 3,334 Patients with Advanced Prostate Cancer Identifies Targetable BRCA Alterations and AR Resistance Mechanisms

Prostate cancer is the second most common cancer in men, accounting for 7% of all cancer-related deaths in this population (1). This cancer is driven by androgen receptor (AR) signaling, and androgen deprivation therapy is the first-line treatment for metastatic prostate cancer. Duration of response varies, with a median of 1–2 years before the disease progresses to castration-resistant prostate cancer (2). AR signaling inhibitors (ARSi) for treatment of metastatic castration-resistant prostate cancer (mCRPC) include abiraterone and enzalutamide, but primary and acquired resistance to these agents remain a challenge (3). Taxane-based chemotherapy has demonstrated efficacy prior and postprogression on abiraterone and enzalutamide (4, 5), yet the 3-year overall survival rate for mCRPC remains under 50% (6).

The landscapes of genomic alterations (GA) of primary prostate cancer (7) and mCRPC (8, 9) have been characterized previously using tissue biopsies and are used to identify mechanisms of resistance to ARSis (10–13). Some GAs enriched in mCRPC are emerging as potential therapeutic targets. Genomic profiling of recent specimens that captures somatic alterations may thus be more valuable than sequencing archival, presystemic treatment primary tumor tissue.

The most recent advances in mCRPC therapy target DNA repair defects in mCRPC using PARP inhibitors (PARPi; refs. 14–16). Rucaparib (17) and olaparib have recently been FDA approved for treatment of mCRPC with germline or somatic BRCA1/2 alterations. Germline alterations in homologous recombination repair (HRR) genes BRCA1/2 are a hereditary risk factor for prostate cancer (18) and mCRPC samples show enrichment in BRCA2 alterations compared with primary tumors, suggesting loss of HRR is a therapeutically relevant driver of aggressive disease (19). Loss-of-function alterations in other DNA damage repair (DDR) genes are also enriched in mCRPC and may be targetable with PARPis, but require further investigation (14–16).

While mutational status of DDR genes can be assessed with smaller sequencing panels, wider panels can detect other alterations targetable in mCRPC, such as PI3K pathway perturbations, which have been targeted with ipatasertib (20), as well as genomic signatures, such as tumor mutational burden and microsatellite instability (MSI), that predict response to immunotherapy and are FDA-approved biomarkers for pembrolizumab in all solid tumors (21, 22).

Despite the inherent advantages of profiling the latest available sample from a patient with advanced disease, mCRPC presents a technical challenge for collection of a tissue specimen, with metastases often confined to bone (23, 24). Bone biopsies are invasive, technically difficult to collect, and have high failure rates of obtaining enough quality DNA for sequencing (10, 25, 26). Blood-based liquid biopsy and genomic profiling of cell-free circulating tumor DNA (ctDNA) from plasma provide a minimally invasive alternate method to profile mCRPC, with the added capability of detecting variants from multiple metastatic lesions that may have undergone clonal evolution. Here, we build off previous studies of ctDNA in mCRPC (27–31) by analyzing a larger cohort of patients, characterizing the genomic landscape leveraging clinically available approaches, and evaluating concordance with tissue-based CGP.

A total of 3,334 liquid biopsy samples and 2,621 tissue samples were assayed with hybrid capture–based comprehensive genomic profiling (CGP). CGP was performed in a Clinical Laboratory Improvement Amendments–certified, College of American Pathologists–accredited, New York State–regulated reference laboratory [Foundation Medicine, Inc. (FMI)]. Patients who submitted screening samples for TRITON2 or TRITON3 provided written informed consent before participation. Approval for the study of the FMI dataset, including a waiver of informed consent and Health Insurance Portability and Accountability Act waiver of authorization, was obtained from the Western Institutional Review Board (protocol 20152817). Studies were conducted in accordance with the Declaration of Helsinki.

Liquid biopsy specimens were obtained from three cohorts: screening samples from the TRITON2 (NCT02952534) and TRITON3 (NCT02975934) trials of rucaparib (collected November 2016–March 2019), and samples submitted to FMI for routine clinical testing (December 2013–March 2019). Cell-free DNA (cfDNA, 20–100 ng) was extracted to create adapted sequencing libraries before hybrid capture and sample-multiplexed sequencing (FoundationACT, FoundationOne Liquid) as described previously (32). Two versions of the plasma assay were used, with 62 (FoundationACT) or 70 genes (FoundationOne Liquid). Genomic regions baited in the two different liquid biopsy assays are depicted in Supplementary Table S1. GAs detected by both assays included base substitutions, insertions and deletions (short variants), rearrangements, and copy-number changes. This study did not evaluate gene deletions. Supplementary Table S2 depicts frequencies of all GAs assessed by liquid biopsy in the three different datasets.

Tissue specimens from metastatic sites submitted for routine clinical testing (December 2013–March 2019) were used for global comparisons of liquid biopsy with metastatic tissue (N = 2,006). Additional tissue specimens used only in the concordance analysis were screening samples collected from the TRITON2 (N = 337) and TRITON3 (N = 277). At least 50 ng of DNA was isolated from prostate cancer acinar adenocarcinoma formalin-fixed, paraffin-embedded tumor specimens and sequenced to high, uniform ≥500× coverage, with larger gene panels inclusive of all 70 genes in liquid assays.

MSI status was determined in samples screened with the 70-gene panel, as described previously (33). For tumor specimens, zygosity and somatic/germline status for mutations were computationally predicted without matched normal tissue, as described previously (34); in validation testing of 480 tumor-only predictions against matched normal specimens, accuracy was 95% for somatic and 99% for germline predictions (35).

Quantification of the ctDNA fraction was measured using two complementary methods: the proprietary tumor fraction estimator (TFE) and the maximum somatic allele frequency (MSAF) method. TFE is based on a measure of tumor aneuploidy that incorporates observed deviations in coverage across the genome for a given sample. Calculated values for this metric are calibrated against a training set based on samples with well-defined tumor fractions to generate an estimate of the tumor fraction. When lack of tumor aneuploidy limits the TFE's ability to return an informative estimate, MSAF is used. MSAF calculates the allele fraction of all known somatic, likely somatic, and variant of unknown significance substitution alterations detected at >2,000× median unique coverage by non-PCR duplicate read pairs, excluding germline variants and variants associated with clonal hematopoiesis.

Liquid biopsy CGP results from a total of 3,334 patients with prostate cancer were included. Patients screened for TRITON3 (n = 818) had progressed on one prior ARSi therapy, while patients screened for TRITON2 (n = 856) progressed on—one to two lines of ARSis, followed by a taxane-based chemotherapy in the castration-resistant setting. A total of 1,660 liquid biopsies were sourced from routine clinical testing at FMI of patients with advanced prostate cancer. The median age of patients was 72 years [range, 38–97; interquartile range (IQR), 66–78] and was similar across all three datasets (Table 1; Supplementary Fig. S1).

Screening samples from the clinical trials contained detectable ctDNA in >95% of samples (as determined by the comprehensive tumor fraction estimation method, see Supplementary Data for full method description). The median ctDNA fraction of these datasets was 7.5% (IQR, 0.8%–34%; Table 1; Fig. 1A) and was higher in patients who had progressed on taxane (median, 18.1% in TRITON2 vs. 3.4% in TRITON3), consistent with higher ctDNA shed after more lines of therapy (29). The number of samples with no ctDNA detected (i.e., no detection of aneuploidy or somatic alleles of known or unknown functional significance) was 40 of 746 (5.4%) in the TRITON3 screening cohort and only 23 of 856 (2.7%) in the more heavily treated TRITON2 screening cohort. A substantial 401 of 856 (47%) samples of the TRITON2 screening cohort had a ctDNA fraction of 20% or above, which allowed for >90% sensitivity of detection for all three variant types reported by the liquid biopsy assay: substitutions/indels, rearrangements, and amplifications. Liquid biopsy identified at least one GA predicted to have deleterious/oncogenic effects on protein function in 2,651 of 3,334 (79.5%) of all patients and 1,888 of 2,213 (85.3%) of patients profiled using the 70-gene panel (Table 1).

The most frequently altered genes were TP53 (46%) and AR (42%; Fig. 1B; Supplementary Fig. S2A), consistent with patient cohorts with prior ARSi exposure; both genes are commonly altered in tissue-based profiling of mCRPC and associated with resistance to ARSis (36, 37). At least one DNA repair gene was altered in 30% of all patients, including BRCA2 (7.5%) and BRCA1 (1.4%). Genes in the PI3K/AKT/mTOR pathway were altered in 14% of samples, including activating mutations in PIK3CA and AKT1 (Supplementary Fig. S3A). WNT/β-catenin pathway genes were altered in 17% of patients (Supplementary Fig. S3B). Alterations in RAS/RAF/MEK pathway components were detected in 5% of all patients (Supplementary Figs. S2A and S3C). Amplifications of FGFR1 were detected in 3% of patients, and 11 patients harbored rearrangements of FGFR1/2/3 with breakpoints in intron 17, which preserve the kinase domain and are predicted to be oncogenic (ref. 38; Supplementary Fig. S3D). MSI-H status was found in 31 of 2,213 (1.4%) of patients (Table 1), comparable with the 2% of primary site and 3% of metastatic tissue biopsies (10, 12).

The landscape of short variants (substitutions and indels) detected by liquid biopsy closely resembled the landscape detected in tissue biopsies from metastatic sites of origin (Fig. 1C; Supplementary Fig. S4A) and previous reports (9, 11, 12). Rearrangements in most genes were detected at similar frequencies in liquid biopsies relative to metastatic tissue (Supplementary Fig. S4B), except AR, as discussed further on. Copy-number amplifications were detected less frequently in liquid than tissue biopsy (Supplementary Fig. S4C), likely owing to decreased sensitivity of detection in samples with low levels of ctDNA (Supplementary Fig. S5). We examined the subset of samples with ≥20% ctDNA fraction (a threshold where there is >95% sensitivity of detection for amplifications; ref. 32). This subset included 1,282 of 3,334 (38%) of all samples (Table 1). Amplifications were detected with significantly higher frequency in this subset: AR (344/781; 41%), FGFR1 (91/1,282; 7.1%), and MYC (67/1,282; 5.2%). It is important to note that the two liquid biopsy platforms used in this study did not report copy-number losses, and common driver TMPRSS2-ERG fusion was not reported because of lack of baiting in these genes.

Frequencies of short variants in liquid biopsies were compared with those in 2,006 metastatic tissue biopsies (Fig. 1C). While most genes were altered at similar rates, variants were detected significantly more often in liquid in 9 of 70 genes. Alterations in AR were enriched in ctDNA, likely representing resistance mechanisms acquired on therapy. Low level enrichment for JAK2, GNAS, and IDH2 (genes not often altered in mCRPC) likely represents signal from clonal hematopoiesis (39), rather than the tumor; the same applies to mutations in NF1 and the TERT promoter. Alterations in CHEK2, ATM, and TP53 occurred with some frequency in mCRPC tissue biopsies, but were also more prevalent among liquid biopsies; it is uncertain whether this is related to mCRPC biology or also to clonal hematopoiesis, as these have been detected in some studies of clonal hematopoiesis (31). In this study, no strong associations with age were observed for likely clonal hematopoiesis variants in JAK2, GNAS, and IDH2 (Supplementary Fig. S6), thus age association could not be used to distinguish clonal hematopoiesis variants from liquid-prevalent resistance mutations, such as AR, sourced from the tumor. The lack of age association for clonal hematopoiesis variants may be the result of an older cohort, in which 88% of patients were older than 60 years.

Patient-matched tissue samples were available for 837 of the 3,334 liquid biopsies (Table 1). Tissue specimens were collected a median of 758 days before plasma collection (range, 19.9 years before to 1.8 years after liquid biopsy; Table 1; Supplementary Fig. S7A). A total of 117 pairs were collected within 30 days of each other and were considered “contemporaneous pairs” in this concordance analysis.

Detection of short variants in genes included in the liquid assay showed 75.3% positive percentage agreement (PPA) to tissue as reference (Supplementary Table S3). PPA was 70.3% for rearrangements and 27.5% for amplifications. Among contemporaneous pairs, PPA increased to 87.2%, 91.7%, and 38.8% for short variants, rearrangements, and amplifications, respectively. ctDNA fraction was a major factor in concordance of liquid to tissue for amplifications because 20% ctDNA fraction is needed to detect amplifications with >95% sensitivity (32); above this threshold, PPA for amplifications was 50.8%. Alterations exclusively detected in tissue tended to be in sample pairs with a low ctDNA fraction in the liquid biopsy (Supplementary Fig. S7B).

BRCA1/2 was altered in 8.9% of all patients' plasma samples (7.5% BRCA2, 1.6% BRCA1, and 0.18% both genes). Curiously, BRCA2 was altered significantly more frequently in the TRITON2 versus TRITON3 screening cohorts (Supplementary Table S4). BRCA2 inactivation is a predictor of poor response to docetaxel (40, 41), and patients screened for TRITON2 had progressed on a taxane-based chemotherapy, which could account for a larger proportion of these patients in this cohort. BRCA1 alterations were also more frequent in the group that had received prior taxane. The combined BRCA1/2 alteration frequency within the TRITON2 screening cohort was 12.3%, nearly twice the frequency in the TRITON3 screening cohort (6.4%; Fig. 2A).

Frameshifts were the most common alteration in both BRCA2 and BRCA1, rearrangements were more common in BRCA1, and the frequencies of nonsense and missense mutations were similar between the two genes (Fig. 2B). A total of 90% of BRCA1/2-altered patients had a single mutation detected (Fig. 2C). In 10 of the 30 patients with two or more mutations, the additional variants were reversions. All 10 patients with BRCA1/2 reversions were from the routine clinical CGP dataset, not the TRITON2/3 screening cohorts (Fig. 2A), and may have had exposure to platinum-based chemotherapy or PARPis (42–44). Reversions were defined as in-frame deletions spanning a frameshift or nonsense mutation, or missense mutations in the same position as the nonsense mutation (Supplementary Table S5). These alterations were mostly subclonal in relation to the deleterious mutation, with up to 13 found in one sample (Fig. 2D).

Studying the variant allele fractions (VAF) of BRCA1 and BRCA2 short variants as compared with ctDNA fraction (Fig. 2E), two distributions were seen, consistent with somatic and germline mutations. Variants with a VAF ≥ 40% detected at low ctDNA fraction were enriched for putative germline variants (i.e., founder mutations). When ctDNA fraction was high, somatic and germline variants were not clearly distinguishable. Some somatic variants tracked closely with ctDNA fractions, while others appeared subclonal (variants close to the horizontal axis in Fig. 2D). The VAF distribution of variants in additional DDR genes, ATM, CHEK2, PALB2, and CDK12, is provided in Supplementary Fig. S8.

Among 837 patients in the concordance analysis, 92 (11%) had a BRCA1/2 alteration detected by one or both assays. A total of 67 (8%) patients had BRCA1/2 alterations detected concordantly in tissue and liquid, five (0.6%) exclusively in tissue biopsy, and 20 (2.4%) exclusively in the liquid biopsy (Fig. 3A and B). The PPA between the tissue and liquid assays was 93.1% on the patient level (Fig. 3B). The variant-level concordance between the tissue and liquid assays for BRCA1/2 was high: the PPA was 95.2% and 85% in the TRITON2 and TRITON3 screening cohorts, respectively (Supplementary Table S3).

Lack of detection of BRCA1/2 variants in liquid was mostly attributable to low ctDNA fraction: of 5 patients with tissue-only detection of a BRCA1/2 alteration, 4 patients had ctDNA fractions of ≤1% (Fig. 3A). Some BRCA1/2 alterations identified exclusively in the liquid biopsy (colored red in Fig. 3A) potentially represent alterations acquired after tissue specimens collection (median time difference in sample collection was 38 months; IQR, 15–112; Supplementary Fig. S7C). These variants, especially in BRCA1, were more likely to be subclonal, have VAFs significantly smaller than the ctDNA fraction of the sample. However, in five of the 20 samples where only liquid biopsy identified BRCA1/2 alterations, BRCA2 short variant VAF exceeded 50% of the ctDNA fraction.

BRCA1/2 alterations detected in tissue were predicted to be germline or somatic, using a previously established and validated computational method (34). All 40 predicted germline mutations were detected in the corresponding liquid biopsies at >40% VAF (Fig. 3C and D). Sixteen of 18 predicted somatic and eight of 10 unknown status were detected in the corresponding liquid biopsy (Fig. 3C), usually at lower VAF (Fig. 3D).

Overall, liquid biopsy was able to detect 100% of predicted germline BRCA1/2 alterations and reliably detect somatic alterations identified in tissue when ctDNA fraction was more than 1%. In samples where BRCA1/2 VAF was lower than ctDNA fraction, it is possible that the variant represents a subclone, a monoallelic alteration, or even clonal hematopoiesis. It is worth noting the assays used in this study detected inactivating short variants and rearrangements, but did not report BRCA1/2 copy loss, thus copy-number deletions were not considered in concordance analysis.

Liquid biopsy detected 1,090 AR alterations among 42% of evaluated patients (940/2,213). Among AR-altered patients, 45% had missense mutations, 33% had amplifications, 8% had rearrangements, and 14% had multiple types of alterations (Fig. 4A). AR amplifications were identified in 419 patients (43% of AR-altered patients and 13% overall), with technically limited detection in samples with low ctDNA fractions (Supplementary Fig. S9). Among the subset of patients with ≥20% ctDNA fraction in their biopsy, AR amplifications were detected in 41% (344/781).

Among patients with a mutation or rearrangement in AR, approximately 40% harbored ≥2 variants (up to six; Fig. 4B; Supplementary Fig. S2B). Many mutations were subclonal, with VAFs smaller than ctDNA fractions; some VAFs exceeded ctDNA fraction, which may result from mutations in amplified copies of AR (Supplementary Fig. S10). Hotspots concentrated in the ligand binding domain (LBD) and are known to confer resistance to ARSis (Fig. 4C; ref. 45). The most frequent mutations included W742L/C (bicalutamide resistance), H875Y, F877L and T878A (bicalutamide/enzalutamide/apalutamide resistance and promiscuous activation by progesterone), and L702H (resistance to abiraterone/enzalutamide, as well as the AR proteolysis-targeting chimera ARV-110, and activation by corticosteroids; refs. 45, 46). Less common AR resistance mutations were found in V716M (15 patients), S889G (15 patients), and M896I/V/L (10 patients; Fig. 4C); these mutations have been detected in ctDNA from patients with mCRPC who progressed on bicalutamide or abiraterone (45, 47, 48).

Rare AR variants were detected in this large dataset, some of which have not been described previously and warrant further characterization. Double mutant F877L/T878A appeared in 11 patients (Fig. 4D). This compound mutant has been shown to confer synergistic resistance to enzalutamide in preclinical studies (45, 49). In-frame deletions spanning residues H875 to T878 were detected in 11 patients, each shifting S885 into the T878 position (Fig. 4D). While it is not known whether these mutants mimic T878S, their appearance in multiple patients in this study raises the possibility of their being a mechanism of ARSi resistance. One patient harbored indel S647–648>F, impacting critical serine residues within a binding motif for the ubiquitin ligase SPOP, and predicted to stabilize the AR by reducing proteasomal degradation (ref. 50; Supplementary Fig. S11A).

AR rearrangements that truncate the reading frame just after exon 3 yield a receptor with an intact DNA binding domain (DBD) and without a LBD to suppress its activity (51, 52). These rearrangements were detected in 160 patients (17% of AR-altered patients and 7% overall), more commonly than the 2.2% frequency detected among metastatic tissue samples (Fig. 4E; ref. 12). The number of intra-AR rearrangements was consistent with ctDNA from patients with mCRPC who have undergone ARSi therapy (52). There were 7 patients with truncating mutations in AR that disrupt the LBD and retain the DBD; these may yield active ARSi-resistant receptors according to similar principles (ref. 53; Supplementary Fig. S11A).

AR alterations were significantly more common among TRITON2 than TRITON3 screening samples (52% vs. 37%; P = 3.0E-8; Supplementary Table S4), consistent with progression on more lines of therapy and higher ctDNA fractions in the TRITON2 cohort. It is also possible that higher ctDNA fractions in the TRITON2 cohort allowed for more sensitive detection of certain variants, such as amplifications.

Among patient-matched tissue/liquid sample pairs, liquid biopsy detected far more AR short variant mutations than tissue (18 shared, 10 tissue only, and 173 liquid only). To exclude pairs where tissue was archival and collected prior to exposure to ARSis, pairs collected within 30 days of each other were compared. Among these, 10 AR mutations were detected concordantly, three mutations exclusively in tissue, and 13 exclusively by liquid. In 7 of 17 patients, liquid biopsy provided the only evidence of AR mutations, and in 2 of 17 patient additional mutations were detected by liquid biopsy (Fig. 4F). This analysis highlights the ability of liquid biopsy to detect subclonal resistance variants that may not exist in every metastatic lesion, and thus, may not be detected in a tissue biopsy, but have relevance when choosing a course of treatment.

AR amplifications were detected in 72 of the tissue samples, with a median copy number of 20 (IQR, 13–35). Liquid biopsy detected 53% of these amplifications (38/72), and 91% of amplifications when ctDNA fraction was ≥20% (32/35). Amplifications with greater copy numbers could be detected at lower ctDNA fractions: liquid biopsy detected 87% of amplifications of copy number > 16 in samples with a ctDNA fraction as low as 5% (Fig. 4G). AR amplifications were detected exclusively in liquid biopsy in 96 samples. Among the 79 contemporaneously collected pairs where AR was profiled (Table 1), 20 amplifications were detected concordantly, 16 exclusively in tissue, and five exclusively in liquid. Lack of detection in liquid biopsy was largely attributable to ctDNA fractions below 20% (Supplementary Fig. S11B). In contemporaneous pairs, one AR rearrangement was detected concordantly and five exclusively in liquid (Supplementary Fig. S11C).

In the largest study of mCRPC liquid biopsy samples conducted to date, CGP of liquid biopsies from 3,334 patients with advanced prostate cancer recapitulated the genomic landscape detected in tissue biopsies, with a high level of agreement in detection of BRCA1/2 mutations. ctDNA also identified more acquired resistance alterations than tissue, including novel AR-activating alterations and subclonal BRCA1/2 secondary mutations and reversions.

Analysis of cfDNA extracted from plasma was feasible in these cohorts, with 94% of patients having detectable ctDNA and with higher ctDNA fraction among patients with more advanced disease. The rich genomic signal within ctDNA, combined with its ease of use as compared with biopsy of a metastatic site, suggest that liquid biopsy could be a compelling option for identifying targetable GAs in patients with mCRPC.

A total of 94% of BRCA1/2 mutations and 90% of all BRCA1/2 variants detected by tissue CGP were detected in ctDNA, including all predicted germline variants. Liquid biopsy also demonstrated the ability to identify patients with somatic BRCA2 mutations not detected in tissue, some with a high VAF, suggesting homologous recombination deficiency acquired later in tumor evolution. Given the high percentage of patients with detectable ctDNA and the agreement with tissue-defined genomic landscape of mCRPC observed in this study, CGP of a minimally invasive plasma biopsy is a viable option to detect BRCA1/2 mutations. Of note, this analysis did not test for BRCA1/2 homozygous deletion; this variant is more challenging to detect in ctDNA, but is measured on the latest version of the assay, which has recently been approved by the FDA (54). An estimated additional approximately 3% of patients with mCRPC might have been identified as having BRCA1/2 mutations with a platform that reports deletions (12). These patients are worth identifying because they may receive particularly sustained benefit from PARPis, having a BRCA defect incapable of reversion (17, 55).

The most striking area of discordance between liquid and tissue variant detection was in the detection of a range of resistance mutations, in agreement with previous reports documenting the sensitivity of liquid biopsies in detecting AR-directed resistance mutations (27, 47, 48, 56–59). AR-activating alterations and subclonal somatic BRCA1/2 secondary mutations and reversions were far more prevalent in liquid biopsies, likely due to a combination of (i) differences in the patient populations, because patients submitting a liquid biopsy likely were exposed to more lines of therapy (Supplementary Fig. S7A), (ii) the increased sensitivity of detection and lower reported VAFs in liquid biopsies over tissue testing, and (iii) the ability of liquid biopsy to integrate acquired resistance signals from multiple metastatic sites (Fig. 4F).

The sensitive detection of resistance mechanisms in liquid biopsies can potentially be used as a marker of therapy resistance and could provide additional ability to detect patients who might benefit from a non-ARSi drug. In this study, liquid biopsy outperformed tissue in detecting AR mutations that have relevance to clinical decisions about choices of ARSis, including W742L (bicalutamide resistance), H875Y and T878A (bicalutamide/enzalutamide/apalutamide resistance), and L702H (resistance to the AR proteolysis-targeting chimera ARV-110; refs. 45, 46). Liquid biopsy's sensitive detection of BRCA1/2 reversions offers potential to be used for monitoring of emerging resistance.

One established limitation of liquid biopsy genotyping is the identification of mutations derived from white blood cells (e.g., clonal hematopoiesis; refs. 29, 31, 60), an age-related phenomenon which may be particularly relevant in patients with prostate cancer who tend to be older at metastatic diagnosis. Intuitively, alterations which confer a fitness advantage in hematopoietic progenitors (JAK2 V617F and IDH2 R140Q) are often suspected to originate from clonal hematopoiesis. Our study found an increased prevalence of mutations in ATM and CHEK2 in ctDNA compared with tissue, two potentially targetable HRR genes that can be mutated in cancer cells, can harbor germline mutations, and have been described as recurrently mutated in clonal hematopoiesis (31, 61). Our study also identified enrichment for NF1 and TERT promoter mutations in ctDNA. NF1 inactivation has been detected in tumor-infiltrating lymphocytes and white blood cells in previous studies (31, 62). TERT promoter mutations have not been implicated as clonal hematopoiesis variants, but have been linked to myeloid malignancies. These findings highlight that genomic discovery based on ctDNA genomics may benefit from paired-depth blood cell sequencing to clarify tumor-derived versus clonal hematopoiesis–derived signal. While such mutations are suspicious for clonal hematopoiesis, even classic clonal hematopoiesis genes may impact prostate cancer biology, with a recent study noting that TET2 and IDH1 are commonly altered in a novel subtype of prostate cancer (63). In clinical care, such paired-depth blood cell sequencing is not widely available, therefore, clinicians must understand, even when using FDA-approved assays, that clonal hematopoiesis mutations are common and mutations in ATM and CHEK2 may not always be tumor derived.

The key limitation of ctDNA analysis is the variable shed of ctDNA into plasma. In this study, 60% of samples had a ctDNA fraction below 20%, which reduced the ability to detect amplifications. Both liquid biopsy platforms used in this study were not designed to detect deletions, leading to marked underdetection of PI3K signaling perturbation (PTEN homozygous deletions account for more than half of alterations in that gene, ∼30% to the 9% detected in this study). Other genes in the panel where deletions were likely missed were RB1, BRCA1/2, APC, and TP53. Deletions are captured with FMI's next-generation assay (54). Finally, the routine clinical CGP prostate cancer samples available for this study do not uniformly represent mCRPC, and there was limited information on treatments the patients received prior to specimen collection. Nevertheless, this dataset closely resembles the defined mCRPC TRITON2/3 cohorts and demonstrates the utility of liquid biopsy to identify patients who may benefit from targeted therapies.

This study highlights the ability of the liquid platform to detect targetable alterations in patients with mCRPC, including somatic mutations acquired during or after the transition to castration-resistant, metastatic disease. It also exposes some of the weaknesses of ctDNA analysis, such as reduced detection of copy-number changes in biopsies with low ctDNA. This reduced sensitivity with lower ctDNA content highlights that liquid biopsy cannot replace tissue CGP, but may complement it. If tissue profiling fails due to specimen inadequacy, the high ctDNA shed in prostate cancer means liquid biopsy could serve as a back-up. Alternatively, if tissue is unavailable for profiling, liquid biopsy could be used first with tissue CGP as a reflex option when ctDNA analysis is negative, indeed clinicians ordering a liquid biopsy might in parallel request archival tissue for CGP so this can be analyzed if ctDNA is uninformative. Together, these two diagnostic tools offer an opportunity to increase access to precision therapeutics in advanced prostate cancer.

H. Tukachinsky reported personal fees from Foundation Medicine and Roche during the conduct of the study. R.W. Madison reported personal fees from Foundation Medicine and other from Roche AG during the conduct of the study. J.H. Chung reported employment with Foundation Medicine. O.V. Gjoerup reported personal fees from Foundation Medicine and other from Roche during the conduct of the study and outside the submitted work. E.A. Severson reported personal fees from Foundation Medicine during the conduct of the study and outside the submitted work. L. Dennis reported other from Roche outside the submitted work. B.J. Fendler reported a patent for WO/2020/236941 pending. S. Morley reported personal fees from Foundation Medicine Inc. during the conduct of the study and outside the submitted work. L. Zhong reported personal fees from Foundation Medicine during the conduct of the study and outside the submitted work. R.P. Graf reported personal fees from Foundation Medicine during the conduct of the study. J.S. Ross reported personal fees from Foundation Medicine during the conduct of the study. B.M. Alexander reported personal fees from Foundation Medicine and Roche during the conduct of the study. W. Abida reported personal fees from Clovis Oncology during the conduct of the study. K. Fizazi reported nonfinancial support from Astellas, Bayer, Janssen, Sanofi, and MSD and personal fees from Clovis, CureVac, and Orion outside the submitted work. S.P. Watkins reported other from Clovis Oncology during the conduct of the study and outside the submitted work. A. Simmons reported other from Clovis Oncology during the conduct of the study. J.M. Venstrom reported other from Foundation Medicine during the conduct of the study, and other from Foundation Medicine and Roche outside the submitted work. G.R. Oxnard reported other from Foundation Medicine and Roche during the conduct of the study and personal fees from AstraZeneca, Takeda, Dropworks, Blueprint, Illumina, Inivata, Janssen, and AbbVie outside the submitted work. No disclosures were reported by the other authors.

H. Tukachinsky: Conceptualization, data curation, formal analysis, investigation, visualization, writing–original draft, writing–review and editing. R.W. Madison: Conceptualization, data curation, formal analysis, investigation, visualization, writing–review and editing. J.H. Chung: Conceptualization, supervision, investigation, writing–review and editing. O.V. Gjoerup: Supervision, project administration, writing–review and editing. E.A. Severson: Resources, supervision, investigation, writing–review and editing. L. Dennis: Supervision, writing–review and editing. B.J. Fendler: Formal analysis, investigation, writing–review and editing. S. Morley: Resources, investigation, writing–review and editing. L. Zhong: Resources, investigation, writing–review and editing. R.P. Graf: Resources, investigation, writing–review and editing. J.S. Ross: Supervision, writing–review and editing. B.M. Alexander: Supervision, writing–review and editing. W. Abida: Resources, investigation, writing–review and editing. S. Chowdhury: Resources, Investigation, writing–review and editing. C.J. Ryan: Resources, investigation, writing–review and editing. K. Fizazi: Investigation, writing–review and editing. T. Golsorkhi: Investigation, writing–review and editing. S.P. Watkins: Resources, Investigation, writing–review and editing. A. Simmons: Resources, investigation, writing–review and editing. A. Loehr: Resources, investigation, writing–review and editing. J.M. Venstrom: Conceptualization, supervision, writing–review and editing. G.R. Oxnard: Conceptualization, supervision, writing–review and editing.

The TRITON2 and TRITON3 trials were funded by Clovis Oncology, Inc. We thank and acknowledge all patients and their families and caregivers who are participating in the trials, along with the investigators. We thank Brady Forcier and Jeff Lee for help with data analysis. We thank Neeru Bhardwaj for critical reading and helpful discussions of the article.

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.