Identification of Hypermutation and Defective Mismatch Repair in ctDNA from Metastatic Prostate Cancer

Metastatic prostate cancer (mPC) has a relatively low mutation rate: approximately four mutations per megabase (Mb) pair (1). Tumor suppressors and oncogenes are frequently altered via structural rearrangements, often manifesting as copy-number changes or gene fusions (2). However, large cohort studies of localized and metastatic cases have identified small subsets of prostate cancer with outlier tumor mutational burden (TMB) and mutations or complex rearrangements within DNA mismatch repair (MMR) genes MSH2, MSH6, or MLH1. Defective MMR (MMRd) is closely linked to increased microsatellite instability (MSI) in other solid cancers (e.g., colorectal), but the relationship between MMRd, high TMB, and MSI is variable in prostate cancer (3, 4). Recent tissue-based studies have identified evidence of MMRd and/or high MSI in 3%–12% of lethal treatment-resistant mPC (1, 3–6).

Predicting patient response to immune checkpoint inhibitors is complex and multifactorial (7), but high TMB is consistently linked with improved response rates (8, 9). It is thought that high TMB leads to expression of more neoantigens and therefore a greater probability of triggering successful T-cell–dependent antitumor responses (10). Prostate cancer has proved challenging to target with immuno-oncology strategies and trials of checkpoint inhibitors in unselected patients have demonstrated very low response rates (11–13). Nevertheless, there are documented MMRd prostate cancers (and therefore, high TMB) exhibiting dramatic responses to immune checkpoint inhibitors (3, 6, 14–16). Clearly, for immunotherapy to be effective in mPC, biomarker-driven selection will be critical.

Plasma circulating tumor DNA (ctDNA) is an established minimally invasive tool for profiling somatic alterations in patients with mPC (17–19). In a recent study we identified high TMB and alterations in MSH2 or MSH6 in two of 115 (1.7%) mPC cases via ctDNA analysis (17). Another group identified high MSI in 3.8% of eligible ctDNA samples collected from patients with mPC (19). Here, we profiled ctDNA from a large cohort of mPC and demonstrate the presence of somatic MMR gene defects and exome-wide signatures of MMRd in 3.7% of patients.

We assembled a meta-cohort of targeted sequencing data from 1,047 plasma cell-free DNA (cfDNA) samples from patients with mPC in an ongoing liquid biopsy program at the University of British Columbia (British Columbia, Canada) and BC Cancer (British Columbia, Canada; refs. 17, 20–22). The vast majority of patients had clinically progressing metastatic castration-resistant prostate cancer (mCRPC) at the time of sample collection, but patients with treatment-naïve metastatic castrate-sensitive prostate cancer were eligible. Samples from this cohort were then included in this study if ctDNA purity as a fraction of total cfDNA was above 2% (17). This threshold was selected because somatic mutations detected in samples with >2% ctDNA show high concordance with matched tumor tissue (20, 23). Approval for collection and genomic profiling of patient samples was granted by the University of British Columbia Research Ethics Board (certificate numbers H18-00944, H14-00738, H16-00934, and H09-01628). The study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all participants prior to enrollment.

Initial targeted sequencing of cfDNA and matched white blood cells (WBC) was performed as previously described using an established prostate cancer gene panel that captures 1.157 Mb of the genome and includes the exonic regions of MSH2, MSH6, and MLH1 (17). Whole-exome sequencing (WES) was performed as described previously (17). CfDNA and WBC sample pairs were subjected to WES if targeted sequencing revealed somatic hypermutation and the cfDNA sample had a ctDNA fraction above 10%. Somatic hypermutation was defined as a mutation rate above the 95th percentile of all ctDNA-positive samples. Targeted and WES data were deposited to European Genome-Phenome Archive under accession number EGAS00001003899.

To assess for the presence of structural rearrangements in samples with evidence of MMRd from initial targeted and WES, we performed an additional round of targeted sequencing using a custom panel that captured the introns and exons of MSH2, MSH6, MLH1, PMS2, and POLE. Probe design and library construction used the SeqCap EZ Choice System (Roche NimbleGen), as per the initial targeted panel.

Patient clinical records were reviewed for availability of archival prostatectomy or diagnostic biopsy specimens. Where possible, formalin-fixed, paraffin-embedded archival tissue was retrieved and submitted for pathology review to identify tumor-rich foci. Tumor tissue DNA extraction and targeted sequencing was performed as previously described using the same initial capture panel as applied to the matched cfDNA (24). Variant calling methodology was consistent across cfDNA and tissue sequencing data.

IHC was performed on 4-μm-thick tissue sections in the clinical laboratory of the BC Cancer Vancouver Centre using the Ventana Discovery XT and the Ventana Benchmark XT Automated System (Ventana Medical Systems) as per the manufacturer's recommendations. Sections were mounted on charged glass slides and baked at 60°C for 15 minutes. Heat-induced antigen retrieval was performed in Cell Conditioning solution CC1 (Ventana). Slides were incubated with anti-hMLH1 antibody (Clone G168-728, Cell Marque), anti-hMSH2 antibody (Clone G219-1129, Cell Marque), anti-hMSH6 antibody (Clone EPR3945, Abcam), and anti-hPMS2 antibody (Clone A16-4, BD Biosciences) for 30 minutes at room temperature. Antibodies were detected using the Ventana DABMap kit, counterstained with hematoxylin, and treated with a proprietary bluing agent (Ventana). Mismatch repair protein expression was considered intact (wild-type) if any percentage of definite positive nuclear staining of the tumor cells was detected. Each protein was considered lost (abnormal) if there was complete loss of nuclear staining in the tumor cells and if there was a positive internal control (intact nuclear staining of benign stromal component such as inflammatory and/or endothelial cells).

Paired-end reads were trimmed to remove adapters, and bases with a quality score <20 were masked. Alignment to the GRCh38.p12 reference genome was performed using BWA-MEM (version 0.7.15; ref. 25); Samtools (version 1.7; ref. 26) was used for sorting and removal of reads with MAPQ < 20. Duplicate reads were marked and removed with Picard (version 2.18.0; Broad Institute). Base level statistics, including depth and nucleotide counts, were generated using pysamstats (https://github.com/alimanfoo/pysamstats).

For targeted sequencing data, somatic single-nucleotide variants and insertion/deletions (indel) were called when variant allele frequency (VAF) was above 1% with at least 10 supporting unique reads in locations with above 30 × depth. Candidate mutations were discarded if VAF was less than 3 × that of the paired WBC or less than 20 × the background error rate, defined as the mean VAF of the substituted base(s) across all WBC controls at the corresponding position(s). For variants adjacent to genomic repeats or regions with a strongly predominant base (>80% of local bases), we required a VAF > 40 × the background error rate. Variants were also filtered if their mean distance from the end of supporting reads was <6 bases. Functional annotation was performed using ANNOVAR (27), and variants identified in the Exome Aggregation Consortium (ExAC) database with population allele frequencies > 5% were discarded. The same methodology was applied to WES data, although a VAF > 2% was required for calling somatic mutations. All final variants from the targeted panels were manually verified in Integrative Genomics Viewer (IGV) 2.3.90 (28).

Putative germline variants were identified from WBC data and defined as nonreference bases with VAF between 30% and 70% and a minimum read depth of 40 ×. The pathogenicity of nontruncating missense variants was assessed using ClinVar annotation of clinical significance and/or the availability of COSMIC identifiers.

A microsatellite was defined as any repetitive DNA sequence identified with RepeatMasker (http://www.repeatmasker.org). Microsatellites were considered unstable if interrupted by an indel, identified using the somatic variant calling criteria above, but only using reads that fully overlapped the microsatellite.

Structural variants were identified using the union of calls from Lumpy (version 0.2.13; ref. 29) and NovoBreak (30) with the requirement that at least 10 split reads supported each breakpoint. All putative rearrangement events that fit these criteria were curated in IGV. The determination of the resulting genotypes was inferred using SSAKE (version 4.0) assembler (31).

Copy-number variants (CNV) were determined from the normalized coverage log ratio across tumor–normal pairs. Read depth of sample pairs was normalized to a low-noise reference by using the median depth of the typically copy-number neutral chromosomes 2, 4, 12, 14, 15, and 19. GC bias was corrected with a Locally Weighted Scatterplot Smoothing (LOWESS) regression. For samples with a ctDNA fraction > 20%, genic CNVs were classified on the basis of mean normalized coverage log ratio. Monoallelic loss or single-copy gains required an absolute log ratio between 0.35 and 0.7, and deep deletions or amplifications required an absolute log ratio > 0.7.

Castration resistance was defined according to Prostate Cancer Clinical Trials Working Group 3 guidelines (32). Time to castration resistance was calculated from the start of androgen deprivation therapy (ADT) for recurrent or advanced prostate cancer. Clinical outcomes for MMRd patients were compared against a subset of 199 patients with mCRPC from within our meta-cohort had outcomes published as part of a randomized phase II study and who showed no evidence of MMRd (17). Genomic comparisons were performed against the subset of this control cohort with above 2% ctDNA (n = 114/199). WES data from 61 high ctDNA samples in the control cohort were used as a comparator for the WES results from the hypermutated cohort. Survival fractions were estimated using the Kaplan–Meier method and differences between groups were estimated using the log rank test. All hypothesis tests were two-tailed and used a 5% significance threshold. Statistical tests and visualizations were performed or generated in Python V3.6 using matplotlib, pandas, numpy, scipy, statsmodels, and Lifelines survival analysis (33).

We analyzed targeted sequencing data from 665 ctDNA-positive plasma samples from 433 patients with mPC (predominantly mCRPC but including 36 patients with castrate-sensitive disease at the time of their first ctDNA-positive blood draw). The mean and median number of detected somatic mutations was 5.3 and 3, respectively (range 1–180; Fig. 1A; Supplementary Table S1). This corresponds to a median somatic mutation rate estimate of 2.6 per Mb, consistent with previous blood and tissue-based estimates in mPC (1, 2, 17, 19).

The top 5% of the cohort exhibited outlier high mutation rates: at least one ctDNA sample from 24 patients had over 11 mutations per Mb (Fig. 1A). Among these 24 hypermutated patients, 10 had either a somatic MSH2 or MSH6 coding mutation or deep deletion (Fig. 1B; Supplementary Figs. S1–S4; Supplementary Tables S2 and S3). Furthermore, 1 patient within the 5% hypermutated subgroup had a germline MSH6 missense mutation (p.P991L) of unknown significance and somatic LOH. Deleterious somatic mismatch repair gene mutations or deep deletions in MSH2, MSH6, or MLH1 were not detected in any patient without a hypermutated sample (10/24 vs. 0/409; P = 3.14 × 10⁻¹⁴; Fisher's exact test).

Of the remaining 13 patients (without a MMR gene coding–region mutation or deep deletion) with an outlier mutation rate, there were 4 patients with deleterious alterations in BRCA2. Two had a germline BRCA2 mutation and somatic LOH, while the others had either a somatic BRCA2 mutation or deep deletion (Fig. 1B). Defective homologous recombination repair (HRR), enabled by BRCA2 loss, results in genome instability including base substitutions as well as widespread large indels (34). Furthermore, in 2 patients, the outlier mutation count from targeted sequencing was potentially explained by localized hypermutation (termed “kataegis;” ref. 35) affecting either the AR gene 3′-untranslated region or a region on chromosome 16.

Notably, there were several cases with distributed hypermutation and no apparent truncal biallelic HRR or MMR defect (Fig. 1B and C; Supplementary Fig. S1). We considered it plausible that some of these samples harbored either defects in MMR genes not covered by our initial targeted panel (e.g., PMS2), or intronic structural rearrangements in MSH2, MSH6, or MLH1 that would remain cryptic to an exome-limited approach (5).

MMRd results in a distinctive pattern of genomic scarring that cannot be easily discerned from targeted sequencing data (36). Furthermore, targeted sequencing can overestimate genome-wide mutation rate (37). Therefore, we performed WES on cfDNA samples (median ctDNA fraction 43.2%) from 21 of the 24 patients with hypermutation from initial targeted sequencing, and compared these data with 61 control mPC samples that did not exhibit an outlier mutational burden from targeted sequencing. Median on-target exome sequencing depth across samples from the hypermutation cohort was 185 × (Supplementary Table S4). Note that 3 of the 24 hypermutated patients (all with MMRd identified by targeted sequencing) could not be subjected to WES due to insufficient ctDNA fraction or library construction failure.

All samples with deleterious protein coding mutations or deep deletions in MSH2 and/or MSH6 demonstrated exome-wide hypermutation (Fig. 2A and B; Supplementary Tables S2 and S4). The median total mutation count in these patients was 1,919, compared with 60 in the control cohort (P = 8.57 × 10⁻⁶; Mann–Whitney U test). The 6 patients with truncal BRCA2 defects and/or kataegis from targeted sequencing showed a significantly lower exome-wide somatic mutation rate than the patients with known MMR gene defects (P = 0.0012; Mann–Whitney U test). Interestingly, a further 6 patients without an MMR gene defect from targeted sequencing also demonstrated exome-wide hypermutation (median mutation count in these patients of 762.5), potentially consistent with MMRd etiology (Fig. 2A and B). None of the control cohort showed exome-wide hypermutation, and overall mutation rate was correlated between targeted and WES (Fig. 2B).

Samples with exome-wide hypermutation in-line with MMRd also showed a high rate of C>T transitions, particularly in the NCG trinucleotide context (Fig. 2C and D). This is consistent with trinucleotide signatures attributed to MMRd (COSMIC signatures 6, 15, and 21; refs. 38, 39). It is also similar to the trinucleotide signature associated with aging (e.g., COSMIC signature 1), a mutational process that is prevalent in most prostate cancers (ref. 40; Supplementary Fig. S5). The median proportion of signatures 6, 15, and 21, was significantly higher in samples with exome-wide hypermutation (with or without an obvious MMR gene defect) compared with the control cohort (0.41 vs. 0.15; P = 0.00068; Mann–Whitney U test; Fig. 2C).

MSI is a continuous variable, and nonarbitrary thresholds for determining high versus low MSI are challenging to derive. Nevertheless, the patients with evidence of MMRd-related hypermutation demonstrated higher MSI than all other samples (median number of unstable microsatellites 56 vs. 1 in the control cohort; P = 2.1 × 10⁻⁸; Mann–Whitney U test; Fig. 2E and F). No samples from the non-hypermutated patients had more than four unstable microsatellites (Fig. 2E and F). Unsurprisingly, we observed a strong relationship between the numbers of unstable microsatellites and indels (r = 0.727 across all samples with WES data; P = 2.22 × 10⁻¹⁶; Spearman correlation; Fig. 2F).

Overall, initial targeted sequencing and WES suggested that 16 patients had hypermutation with MMRd etiology (3.7% of the ctDNA-positive mPC cohort; 16/433), including 6 patients who did not have an obvious somatic MMR gene mutation or deletion. Therefore, to identify MMR gene defects that would be undetected by exome-only sequencing, we performed an additional round of targeted sequencing across the full genic regions of MSH2, MSH6, MLH1, PMS2, and POLE. These data confirmed all prior MMR gene coding–region mutations, and also provided the underlying MSH2 intronic breakpoints that resulted in biallelic deletion of neighboring gene MSH6 in P04 (Supplementary Fig. S6). In addition, we identified a MSH6 intronic rearrangement in P07, as well as a 450-bp deletion encompassing exon 7 of MLH1 in P10 (Supplementary Fig. S6). Both of these patients had MMRd etiology by WES analysis (Fig. 2). Note that no deleterious alterations in POLE or PMS2 were identified via this approach or WES.

Seven patients had appropriate archival tumor tissue for IHC of MMR proteins, and results supported the cfDNA sequencing classification in all patients (Supplementary Fig. S7). The two cases with either BRCA2 loss or kataegis (and no evidence of MMRd by WES) showed wild-type MMR protein expression, while patient P01 (deep deletion of MSH2/6 loci) showed no MSH2 and MSH6 expression, and P18 (MSH6 mutation) showed attenuated MSH6 expression. In support of the MLH1 structural rearrangement identified through MMR gene intron sequencing, P10 showed no protein expression of MLH1. P20 and P05, who had MMRd etiology by cfDNA sequencing (but no causative defect identified), had no protein expression of MSH2 and/or MSH6 (Supplementary Fig. S7): nominating cryptic genomic or epigenetic defects in these genes as the underlying mechanism of MMRd.

Samples with MMRd had a higher rate of mutation and lower rate of copy number alterations in most genes within our targeted panel, compared with MMR-intact mPC (Fig. 3A; Supplementary Tables S5 and S6). For example, TP53 was mutated in all except 3 patients, while the AR gene (typically mutated in only 10%–20% of mCRPC) showed hotspot missense mutations in at least one sample from 9 of 16 patients (with 6 patients exhibiting at least two different AR mutations; Fig. 3B; Supplementary Table S7). The AR mutation prevalence is significantly higher than MMR-intact mPC cfDNA analyzed with the identical approach (9/16 vs. 12/114; P = 7.8 × 10⁻⁵; Fisher's exact test). PTEN is typically altered by copy-number loss in mPC and mutated in only 5% of MMR-intact cases, but 8 of 16 MMRd cases had at least one PTEN protein–coding mutation. Of particular interest, because activating point mutations in oncogenes (other than the AR) are very rare in mPC, hotspot mutations in genes such as PIK3CA (n = 3), AKT1 (n = 4), and CTNNB1 (n = 6), were common in MMRd prostate cancer (Supplementary Table S7).

Individual tumor suppressor genes frequently harbored more than two mutations within a sample (Fig. 3A). Although the short fragment size of cfDNA prevents phasing of most cooccurring mutations (to determine whether they fall on different alleles), differences in mutation allele frequency suggested heterogeneity in their clonal distribution (Supplementary Fig. S8). We defined putative subclonal mutations as those with a VAF less than one quarter of the estimated ctDNA fraction. Consistent with the concept of highly polyclonal disease, all MMRd patients had at least one sample which displayed subclonal mutations (average proportion of subclonal mutations in samples = 0.39; range 0–0.77; Supplementary Table S8). While the underlying MMRd predisposes to continual accrual of passenger mutations, in many patients we observed subclonal oncogene hotspot missense mutations (Fig. 3C; Supplementary Tables S6–S9). This suggests that some of the low VAF mutations may represent emerging clones with different fitness properties to the predominant tumor clone.

Nine hypermutated MMRd patients provided more than one cfDNA sample during their disease course (Fig. 3C). Each sample was collected at progression on a different systemic therapy (predominantly AR pathway inhibitors enzalutamide or abiraterone), and represents treatment-resistant disease. Although most patient-specific mutations were detected in all time points, we observed a highly dynamic temporal landscape, underscored by differences in mutation VAF between time points (Fig. 3C). This included the emergence and disappearance of several AR ligand–binding domain mutations (Supplementary Table S7). These data again suggest the presence of multiple subclonal populations in flux.

Twelve patients, including 9 MMRd, had archival primary tissue samples (predominantly diagnostic needle biopsy) available for targeted DNA sequencing. Somatic mutation counts between same patient tissue and ctDNA were correlated and all except 2 MMRd patients exhibited a degree of hypermutation in their primary tissue (Supplementary Figs. S9 and S10). In these 2 patients (P02 and P18), biopsy tissue represented Gleason Grade Group (GGG) 1 and 2/3 cancer, respectively, and likely undersampled disease that was metastatic at the time of diagnosis. Overall, there was variability in the presence of driver gene alterations in archival tissue compared with ctDNA (Supplementary Figs. S10 and S11). For example, in patient P09, analysis of a single diagnostic biopsy core identified 45 somatic variants, 28 of which (62%) were not detected in any of the subsequent ctDNA samples despite VAFs in the tissue up to 20.9% (Supplementary Fig. S11). Conversely, while the predominant tumor clone in the archival untreated tissue was characterized by a FOXA1 forkhead domain mutation, the predominant clone(s) in the castration-resistant setting also harbored multiple AR and TP53 mutations (Supplementary Table S7; Supplementary Fig. S8).

Recognizing that analysis of a single biopsy core can undersample primary disease, in patients P04 and P10 we analyzed multiple regions of their archival primary tumor tissue (obtained via prostatectomy and tissue biopsy, respectively), using the same targeted sequencing approach as applied to cfDNA samples. Again, we observed high variability between tumor tissue and subsequent ctDNA collections, but also between matched tissue samples (Fig. 4; Supplementary Table S10).

Across both patients, 24% of mutations detected in the tissue were not subsequently detected in cfDNA. Furthermore, only 64% of mutations were shared between any two matched tissue pairs, and only 28% and 14% of somatic mutations were shared by all tissue samples in patient P04 and P10, respectively (Fig. 4; Supplementary Table S10). Because this suggests high clonal diversity, even within the primary tumor, we constructed maximum parsimony trees to illustrate probable phylogenetic relationships between tissue samples (Fig. 4). This revealed distinct ancestral clades, marked by prostate cancer driver mutations. In both patients, there was a clear common ancestor for all tissue and ctDNA samples, and evidence that a particular clade within the primary tissue had preferentially contributed to metastatic spread (and was subsequently represented in ctDNA). In patient P04, there appeared to be a clade that was largely absent from ctDNA at the time of CRPC progression (Fig. 4A). This clade was characterized in part by an unusual AR ligand–binding domain mutation p.V716M. Because patient P04 had undergone radical prostatectomy, it is plausible that surgery removed the constituent clones of this clade from the patient.

Conversely, in P04 and P10, 31 and 17 mutations were restricted to one or more cfDNA sample (i.e., not detected in any tissue sample). These included hotspot AR mutations in both patients, likely selected by ADT, and TP53 mutations in P10 (Fig. 4). Together, these data suggest that a single primary tissue sample is not representative of posttreatment metastatic disease in MMRd patients, potentially due to the dynamic clonal landscape enabled by continual accrual of new mutations.

Eleven of the 16 MMRd patients had evaluable clinical data. Patient characteristics are summarized in Table 1 and Supplementary Table S11. MMRd patients had features consistent with aggressive disease at diagnosis and displayed a significantly shorter time from commencement of ADT to CRPC progression than the control cohort of 199 patients with mPC (Table 1; Fig. 5A). Five patients (50%) achieved a greater than 50% PSA decline from baseline (PSA50) with an AR pathway inhibitor, compared with 70% in the control cohort (Fig. 5B; Supplementary Table S11). The median duration of first-line therapy was only 3.9 months (range 0.9–13.9). Overall, the MMRd patients identified here had a shorter overall survival with CRPC than the control cohort (Fig. 5C). Four patients received only one line of therapy (Supplementary Fig. S12). There was no PSA50 to taxane chemotherapy given in the castration-resistant setting among the 8 MMRd patients in this study.

We demonstrate that somatic hypermutation and presence of an underlying MMR gene defect can be inferred from sequencing of plasma cfDNA, in patients with mPC with abundant ctDNA. This is consistent with results from Mayrhofer and colleagues who showed that detection of high MSI in ctDNA is associated with elevated mutation load (19). Our work suggests that a comprehensive dissection of mutation load, distribution, and signature, in addition to MSI assessment, is required to identify hypermutated cases with cryptic MMR gene defects.

Overall, we report a similar prevalence of MMRd (3.7% vs. 4.5%) to a recent tissue-based characterization that included 356 mCRPC cases (3). However, other tissue-based studies have reported evidence for MMRd in up to 12% of mCRPC cases (1, 4–6). A caveat of our result is that patients without evidence of ctDNA in their plasma were excluded from study. Nevertheless, it is clear that MMRd mPC is a rare genotype. Among the cases in our study where archival diagnostic tissue was available, evidence of MMRd etiology was generally present in the primary tumor, although there were two exceptions (in patients where the biopsy appeared to have undersampled aggressive disease). Abida and colleagues reported several cases where an MMR gene defect appeared to arise late during disease progression, suggesting that sampling of archival primary tissue rather than contemporary metastatic lesions may mischaracterize MMRd status (3). Interestingly, in our study, although most samples had relatively few copy-number changes, some demonstrated copy number features that are thought to be early or initiating events in prostate cancer, such as NKX3-1 deletion (41). This supports the hypothesis that MMRd prostate cancer can arise secondarily from a clone of conventional structural rearrangement-driven adenocarcinoma.

Prior work has established that TMB is correlated between targeted sequencing and WES (42). Patients with diffuse genomic hypermutation tend to have outlier mutation counts in any given subset of the genome, and are therefore highly likely to have hypermutation by targeted sequencing. However, as our data demonstrate, detection of apparent somatic hypermutation from targeted sequencing is not specific to MMRd in prostate cancer. Indeed, targeted sequencing overestimates global TMB because of a focus on genes that are under selective pressure to accrue mutations. Furthermore, underlying MMR gene alterations can be cryptic or epigenetic in nature, and therefore not easily identified by genomic analysis. While our data suggests they can be inferred by assessment of TMB, MSI, and trinucleotide mutational signatures (TNS), small-panel sequencing of either tumor tissue or ctDNA does not typically provide enough data points for these analyses. In future, sequencing-based strategies aimed at specifically identifying hypermutated MMRd prostate cancer may need to employ WES or whole-genome sequencing. Nevertheless, our results suggest that cfDNA sequencing represents a minimally invasive practical tool for this type of analysis.

Patients in our study with detectable ctDNA and hypermutated MMRd etiology responded poorly to AR pathway inhibition, the mainstay of mPC therapy. A key limitation of our approach is that it may exaggerate prognostic trends because detection of ctDNA is itself a poor prognostic factor (43). Our results are broadly consistent with the conclusions of two other studies that leveraged tissue-based profiling to identify a similarly small proportion of MMRd cases (3, 4), but they are in contrast to a third tissue study where patients with MMRd mPC derived benefit from AR pathway inhibitors (ref. 6; Table 1). It is therefore not appropriate to assume that all MMRd mPC will respond poorly to AR pathway inhibition. In all likelihood, MMRd mPC aggression is a factor of patient population, lead time, and the precise pattern of genomic alterations enabled by the DNA damage repair defect.

Our data nominates several potential explanations for poor responses to AR pathway inhibition in our cohort. First, we observed high intrapatient heterogeneity including evidence of branching subclonal tumor populations that fluctuated over time. This diverse ecosystem of tumor clones may increase the probability that any one clone has primary resistance to a given therapy. Second, we saw widespread oncogene activation via hotspot missense mutation. This is rare in conventional mPC and may indicate a degree of oncogenic redundancy and less reliance on the AR to enable proliferative cellular programs. Third, our cases were enriched for AR ligand–binding domain mutations and displayed propensity for rapid switching of AR mutation genotype. AR mutations can render AR inhibitors ineffective or even agonistic, resulting in disease progression (44). Importantly, however, because it remains unclear whether all MMRd cases respond poorly to AR pathway inhibition, these correlative hypotheses must be tested in prospective studies.

Taxane-based chemotherapy is an alternative for mCRPCs with AR alterations that are unlikely to respond to AR pathway inhibitors (45, 46). There was no clinically meaningful PSA response to chemotherapy among MMRd patients in our study, although chemotherapy was used mostly in the second-line setting or beyond, and often as “salvage therapy” in rapidly progressing disease. Conversely, in two recent studies, approximately 50% of high MSI (MSI-H)/MMRd patients (total n = 15) achieved a PSA50 decline during immune checkpoint inhibitor treatment (3, 6). Although patient numbers in these studies were small, the result is consistent with the post hoc analysis of CheckMate 650, where patients with mCRPC with elevated TMB treated with combination of nivolumab and ipilimumab experienced significantly longer radiological progression-free survival (PFS), and higher objective and PSA response (47). Combined with our data, this nominates the hypothesis that ctDNA analysis can help select patients with mPC with the greatest probability of responding to immune checkpoint inhibitors.

A key limitation of a ctDNA-based approach is that tumor material in the blood represents an average of all contributing tumor populations. While this has the theoretical advantage of capturing intrapatient tumor heterogeneity, it is very challenging (and in some cases impossible) to resolve clonal structures. Future studies should profile circulating tumor cells or metastatic tissue biopsies in combination with ctDNA to provide greater insight into specific subclones.

S.Y.F. Fu reports receiving other remuneration from Roche. I. Kushnir holds ownership interest (including patents) in Teva. K.N. Chi is an employee/paid consultant for Astellas, Janssen, Bristol-Myers Squibb, Sanofi, Bayer, and Roche, and reports receiving commercial research grants from Astellas, Janssen, and Sanofi. A.W. Wyatt reports receiving commercial research grants and speakers bureau honoraria from Janssen. No potential conflicts of interest were disclosed by the other authors.

Conception and design: E. Ritch, M. Annala, K.N. Chi, A.W. Wyatt

Development of methodology: E. Ritch, S.Y.F. Fu, C. Herberts, K.N. Chi, A.W. Wyatt

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S.Y.F. Fu, G. Wang, E.W. Warner, E. Schönlau, G. Vandekerkhove, K. Beja, Y. Loktionova, D. Khalaf, I. Kushnir, C. Ferrario, S. Hotte, K.N. Chi

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E. Ritch, S.Y.F. Fu, C. Herberts, E.W. Warner, S. Taavitsainen, A.J. Murtha, M. Annala, K.N. Chi

Writing, review, and/or revision of the manuscript: E. Ritch, S.Y.F. Fu, C. Herberts, G. Wang, I. Kushnir, C. Ferrario, S. Hotte, K.N. Chi, A.W. Wyatt

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E. Ritch, S.Y.F. Fu, C. Herberts, E. Schönlau, K. Beja, S. Hotte, K.N. Chi

Study supervision: K.N. Chi, A.W. Wyatt

Other (pathology): L. Fazli

This work was supported by a Canadian Institutes of Health Research project grant, the Prostate Cancer Foundation, and Prostate Cancer Canada (all to K.N. Chi and A.W. Wyatt).

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.