Genomic alterations in cell-free DNA and enzalutamide resistance in castration-resistant prostate cancer

Abstract

Importance

The molecular landscape underpinning response to the androgen receptor (AR) antagonist enzalutamide in metastatic castration-resistant prostate cancer (mCRPC) patients is undefined. Consequently, there is an urgent need for practical biomarkers to guide therapy selection and understand resistance. Although tissue biopsies are impractical to perform routinely in the majority of mCRPC patients, the analysis of plasma cell-free DNA (cfDNA) has recently emerged as a minimally-invasive method to explore tumor characteristics.

Objective

To reveal genomic characteristics from cfDNA associated with clinical outcomes on enzalutamide.

Design

We collected temporal plasma samples (baseline, 12-week, end-of-treatment) for circulating cfDNA and performed aCGH copy number profiling and deep AR gene sequencing. Samples collected at end-of-treatment were also subjected to targeted sequencing of 19 prostate cancer associated genes.

Setting

Plasma samples were obtained from August 2013 to July 2015 at a single academic institution (British Columbia Cancer Agency).

Participants

65 mCRPC patients.

Exposure for Observational Studies

Enzalutamide, 160 mg daily orally.

Main Outcome Measures

PSA response rate (decline ≥ 50% from baseline confirmed ≥ 3 weeks later). Radiographic (as per Prostate Cancer Working Group 2 Criteria) and/or clinical progression (defined as worsening disease-related symptoms necessitating a change in anti-cancer therapy and/or deterioration in ECOG performance status ≥ 2 levels).

Results

cfDNA was isolated from 122/125 plasma samples, and targeted sequencing was successful in 119/122. AR mutations and/or copy number alterations were robustly detected in 46% and 66% of baseline and progression samples respectively. Detection of AR amplification was associated with primary resistance, as was heavily mutated AR (≥2 mutations), and RB1 loss. AR mutations exhibited clonal selection during treatment, including an increase in glucocorticoid-sensitive AR-L702H and promiscuous AR-T878A in patients with prior exposure to abiraterone. At the time of progression cfDNA sequencing revealed mutations or copy number changes in all patients tested, including clinically-actionable alterations in DNA damage repair genes and PI3K pathway genes, and a high frequency of activating CTNNB1 mutations.

Conclusions and Relevance

These data demonstrate that clinically-informative genomic profiling of cfDNA is feasible in nearly all mCRPC patients and provides important insights into enzalutamide response and resistance.

Introduction

Prostate cancer cells are initially reliant on circulating androgens binding and activating endogenous Androgen Receptor (AR). Although androgen-deprivation therapy elicits a response in most patients, progression to castration-resistant prostate cancer (CRPC), driven frequently by AR reactivation, is inevitable¹. However, in recent years, continued targeting of the AR signaling axis with abiraterone acetate (abiraterone) and enzalutamide, have changed clinical practice and improved the overall survival of CRPC patients^2–5. Multiple other AR axis inhibitors are in clinical trials, promising to expand the arsenal available to patients⁶.

Despite the efficacy of abiraterone and enzalutamide, significant challenges persist. Resistance is inevitable, including primary resistance in 10% of chemotherapy-naïve patients². Although abiraterone and enzalutamide act on different aspects of AR signaling (abiraterone is an androgen synthesis inhibitor; enzalutamide is an AR ligand binding domain (LBD) antagonist), cross-resistance is common^7,8, and likely to be further compounded by the introduction of additional analogous agents. Resistance mechanisms are emerging, and although some tumors evolve to become ‘non-AR driven’, the majority seem to yet again reactivate the AR through heterogeneous genomic and transcriptomic alterations, similar to front-line CRPC^6,9.

There is an urgent need for practical biomarkers to guide therapy selection and understand resistance. Although tissue biopsies have proved informative, they are impractical to perform routinely in clinical practice due to the bone-predominant metastatic landscape of CRPC, and focus has turned to the development of minimally-invasive biomarkers from blood or urine^10–13. Already, the profiling of circulating tumor cells (CTCs) and cell-free DNA (cfDNA) has revealed that truncated AR splice variants, AR copy number gain and point mutations of AR are linked to resistance to abiraterone^12–15. The genomics of enzalutamide response is less well-studied, although AR splice variants are associated with resistance¹², and we recently demonstrated that the presence of certain AR alterations in baseline cfDNA was a poor prognostic factor¹⁴. However, we did not profile the full spectrum of AR mutations, nor did we examine temporal samples to determine changes at progression. Therefore, we hypothesized that comprehensive genomic profiling of sequential cfDNA samples from a larger cohort of CRPC patients treated with enzalutamide would reveal novel molecular characteristics associated with clinical outcomes.

Methods

Patient cohort

Plasma was collected at baseline, 12-weeks and end-of-treatment from 65 patients with metastatic CRPC (mCRPC) who received enzalutamide (see study design in eFigure 1 in Supplement 1). All patients were recruited from British Columbia Cancer Agency, Vancouver, Canada (with the approval of the University of British Columbia Research Ethics Board (Protocol/Approval #H13-00870) and continued on enzalutamide until clinical or radiographic progression. Written informed consent was obtained from all participants prior to enrolment. Baseline patient characteristics are presented in Table 1 (and eTable 1 in Supplement 2). Progression was classified as radiological (as defined by Prostate Cancer Working Group 2 criteria)¹⁶ or clinical (defined as worsening disease-related symptoms necessitating a change in anti-cancer therapy and/or deterioration in ECOG performance status ≥2 levels)¹⁷. cfDNA extraction, processing and array Comparative Genomic Hybridization (aCGH) were performed as previously described although formalin was not added at blood collection¹⁴.

Table 1.

Baseline patient characteristics.

Characteristic (n = 65)
Age
Median (IQR)	74 (68–79)
Gleason Score, n (%)
6–7	22 (34)
8–10	36 (55)
Unknown	7 (11)
Disease sites, n (%)
Bone	62 (95)
Lymph nodes	29 (45)
Visceral	9 (14)
ECOG PS, n (%)
0–1	47 (72)
≥2	18 (28)
Laboratory
Hemoglobin (median, g/L) (IQR)	122 (108–128)
≤100, n (%)	11 (17)
LDH (median, U/L) (IQR)	236 (190–330)
Elevated, n (%)	22 (34)
ALP (median, U/L) (IQR)	130 (97–242)
Elevated, n (%)	28 (43)
Albumin (median, g/L) (IQR)	39 (36–41)
Low, n (%)	12 (24)
All treatments received
Abiraterone acetate	40 (62)
Docetaxel	27 (42)
Bicalutamide	60 (92)
Flutamide	11 (17)
Nilutamide	15 (23)
Other	20 (31)

IQR: Interquartile range, ECOG PS: Eastern Cooperative Group Performance Status, LDH: Lactate Dehydrogenase, ALP: Alkaline phosphatase

cfDNA sequencing

For deep AR sequencing, we used PCR to amplify fifteen ~150bp fragments coding for AR exons 2–8, in three pools of PCR reactions each using 1ng of cfDNA. Universal adapters and barcodes were added by PCR and libraries were sequenced using an Illumina MiSeq. For targeted sequencing of prostate cancer genes we used an Ion Ampliseq Custom DNA panel (Life Technologies) to capture the exons of 19 genes (AR, ASXL1, BRCA1, BRCA2, CHD1, CHEK2, CTNNB1, FOXA1, HSDB31, KDM6A, MED12, KMT2A, MYC, OR5A1, PIK3CA, PTEN, SCN11A, SPOP and TP53). For detailed protocols and description of data analysis see eMethods in Supplement 1. All patient cfDNA sequencing reads are available at the European Nucleotide Archive, accession number PRJEB11659.

Statistical analysis

Univariate analysis was performed to identify variables associated with PSA response (χ² for categorical variables or logistic regression for continuous variables) and clinical and/or radiographic PFS (Cox proportional-hazards modeling). Survival functions were estimated using the Kaplan-Meier method and compared using the log-rank test.

Results

Patient cohort and approach for deriving AR mutation status from cfDNA

In the 65 mCRPC patients recruited to our study, PSA50 (PSA decline ≥ 50% from baseline confirmed ≥ 3 weeks later) and PSA30 response rates to enzalutamide were 38% (25/65) and 50% (31/62) respectively, while median clinical/radiographic progression-free survival was 3.5 months (95% CI 2.1–5.0). Baseline plasma samples were collected for all patients and for 30/65 patients we collected a second plasma sample at 12 weeks post-baseline, prior to progression. Finally, we collected an end-of-treatment plasma sample from 30/65 patients at the time of progression on enzalutamide. Extraction of cfDNA was successful from 122/125 samples; the median DNA concentration was 1.39ng/µl (20–60µl elution; eFigure 2 in Supplement 1).

Each cfDNA sample was split, with half undergoing aCGH¹⁴, and the remaining unamplified cfDNA reserved for next-generation sequencing (NGS). Tumor-derived cfDNA (known as circulating tumor DNA or ctDNA) is typically highly degraded, diluted (with cfDNA from non-cancerous cells), and present in low amounts. Therefore, mutational profiling of all CRPC patients necessitates a sensitive approach that can resolve low proportions of ctDNA among the total cfDNA burden. Given the importance of the AR in CRPC, we designed an assay to sequence the AR gene exons 2–8 (including the commonly mutated LBD) from 1–3ng of cfDNA (eFigure 3 in Supplement 1). In a calibration experiment with prostate cancer cell lines 22RV1 and LNCaP, endogenous AR H875Y and T878A mutations were robustly detected at frequencies as low as 1–2% (eFigure 4 in Supplement 1). Deep AR sequencing was successful in 119/122 patient samples (98%), with an average median coverage >31,000×.

Genomic changes in baseline cfDNA are associated with disease progression

By aCGH, 27/63 (43%) baseline cfDNA samples harbored detectable changes in genome copy number (eTable 2 in Supplement 2). The remaining 36 samples with no detected changes likely harbor a ctDNA proportion below the detection limit of aCGH. Aberrations were typical of prostate cancer, including chr8p loss (15/63), chr8q gain (21/63), and RB1 loss (13/63) (Figure 1A; eFigure 5 in Supplement 1). AR copy number gain was detected in 19 patients, with 10/19 demonstrating evidence of amplification (logratio >1.2).

Figure 1.

AR mutations detected in the cfDNA of mCRPC patients treated with enzalutamide. a) The most common aberrations detected by aCGH and deep AR gene sequencing of cfDNA from patients at baseline. b) Landscape of AR gene mutations detected in the cfDNA of mCRPC patients at baseline and at progression on enzalutamide. Schematic of the AR protein domains mapped onto the exonic structure, showing the pile-up of recurrent mutations detected in the ligand binding domain. Each colored circle represents a single mutation detected in a single sample (i.e. one patient may be represented several times if they harbor multiple mutations or the same mutation in multiple temporal samples). c) Box plot showing the detected allelic frequency of each recurrent mutation baseline and at progression, demonstrating that the majority of patients with AR mutations have at least 5% tumor-derived cfDNA (known as ctDNA). d) High correlation in AR mutation frequency between patient cfDNA technical replicates. Note that the three instances where the validation experiments did not redetect a previously identified mutation, it had a very low original frequency. In those examples, failure of validation likely reflects the low sampling probability of a heavily diluted or rare mutant allele.

We detected AR mutations in 14/62 (23%) patients at baseline on enzalutamide, with a surprising 5/14 having ≥2 mutations (Figure 1A, eTable 2 in Supplement 2). 7 of these 14 patients had no copy number changes detected by aCGH, reflecting the higher sensitivity of the sequencing assay in detecting ctDNA. Notably, no AR mutations were detected in any samples harboring AR amplification (logratio >1.2). This is consistent with recent data that indicated a tendency toward mutual exclusivity¹⁸, and suggests that amplifications and mutations are independent drivers of CRPC progression. The most common mutations at baseline were L702H (6 patients); H875Y (6); T878A (5); and W742L/C (2) (Figure 1B), all known to confer broadened ligand specificity¹⁹. No novel mutations were recurrent. The median allelic frequency of AR mutations at baseline was remarkably high at 8.74% (Figure 1C). Indeed, the allelic frequency of AR mutation was >25% in several samples, suggesting that in some CRPC patients the tumor-derived component of cfDNA is in line with that normally expected of tissue biopsies. Repeating the AR sequencing assay on >30 samples that tested positive for mutations (including those detected at end-of-treatment), demonstrated a high correlation between the original and validation mutation frequency (r² = 0.90; Figure 1D). Finally, since the reported incidence of AR mutations in mCRPC tissue biopsies is 10–20%^20,21, our assay likely captured the vast majority of clinically-relevant mutations.

The detection of AR gain/amplification, multiple AR mutations, RB1 loss, MET gain and MYC gain was linked to adverse outcomes on enzalutamide (Figure 2, Table 2). A potential confounding factor is that these aberrations may simply be surrogate markers for ctDNA detection, which itself was linked to poor outcomes on enzalutamide (Table 2). Nevertheless, after adjusting for the presence of ctDNA, detection of RB1 loss (p=0.01) and MET gain (p=0.02) remained significantly associated with worse progression-free survival (PFS) (Cox proportional-hazards modeling). In addition, there was a strong trend towards shorter PFS for patients with multiple AR mutations (p=0.09). Furthermore, among patients with AR aberrations, those with a ‘heavily aberrant AR’ (amplification or ≥2 mutations) had median PFS of only 1.9 months compared to 4.4 months for patients with a single AR mutation (p=0.035; log-rank). Importantly, patients with a single AR mutation did not appear to exhibit primary resistance to enzalutamide (Figure 2).

Figure 2.

Association of AR gene status at baseline with progression-free survival on enzalutamide. Swimmers plot showing selected copy number changes and AR mutation status in each patient at enzalutamide baseline, as determined by aCGH and deep AR sequencing respectively. *bona fide AR amplification call (log₂ratio > 1.2) only possible in certain patients.

Table 2.

Univariate analysis examining association between baseline genomic aberrations, PSA response (decline ≥ 50% confirmed ≥ 3 weeks later) and radiographic/clinical progression-free survival (PFS) in patients treated with enzalutamide.

	PSA response		PFS
Variable	Frequency	P value	HR	95% CI	P value
MET gain/amp (yes vs. no)	0% vs. 44%	0.02	4.53	1.97, 10.45	<0.001*
RB1 loss (yes vs. no)	8% vs. 46%	0.01	4.46	2.28, 8.74	<0.001*
Tumor-derived cfDNA (yes vs. no)	23% vs. 62%	0.002	4.35	2.15, 8.79	<0.001
Multiple AR mutations (yes vs. no)	20% vs. 39%	0.64	3.94	1.46, 10.64	0.007
Prior abiraterone (yes vs. no)	25% vs. 60%	0.005	3.92	1.99, 7.71	<0.001
AR gain/amp (yes vs. no)	16% vs. 48%	0.02	2.92	1.59, 5.37	0.001
AR amp (yes vs. no)	10% vs. 43%	0.051	2.72	1.30, 5.69	0.008
Serum ALP (> ULN vs. ≤ ULN)	32% vs. 43%	0.36	2.72	1.50, 4.94	0.001
MYC gain/amp (yes vs. no)	22% vs. 44%	0.10	2.58	1.39, 4.77	0.003
Serum LDH (> ULN vs. ≤ ULN)	23% vs. 47%	0.06	2.46	1.36, 4.45	0.03
Visceral metastases (yes vs. no)	22% vs. 41%	0.28	2.43	1.16, 5.07	0.03
ECOG PS (2 vs. 0–1)	28% vs. 43%	0.27	2.38	1.28, 4.43	0.006
Prior docetaxel (yes vs. no)	39% vs. 39%	1.00	2.21	1.19, 4.08	0.01
AR mutation (yes vs. no)	23% vs. 41%	0.24	1.81	0.94, 3.51	0.08
LN metastases (present vs. absent)	38% vs. 39%	0.94	1.59	0.89, 2.86	0.12
Gleason score (≤ 7 vs. ≥ 8)	37% vs. 35%	0.86	1.26	0.69, 2.33	0.45
DNA concentration (continuous)	-	0.28	1.03	1.004, 1.05	0.02
Serum haemoglobin (continuous)	-	0.24	0.98	0.96, 1.01	0.16

AR: androgen receptor, Amp: amplification, Multiple AR mutations: ≥ 2, mCRPC: metastatic castration-resistant prostate cancer, cfDNA: circulating (plasma) cell-free DNA, Tumor-derived cfDNA (ctDNA): PCa-associated copy number change and/or AR mutation, ECOG PS: Eastern Cooperative Group Performance Status, LDH: Lactate dehydrogenase, ALP: Alkaline phosphatase, LN: Lymph node.

^*RB1 loss and MET gain/amp remained significantly associated with shorter PFS after adjusting for presence of tumor-derived cfDNA.

Dynamic cfDNA changes and acquired enzalutamide resistance

Plasma collection from patients at 12-weeks and at end-of-treatment allowed an exploration of factors linked to acquired enzalutamide resistance. While undergoing treatment, only 8/29 (28%) samples had evidence of ctDNA (eTable 2 in Supplement 2), consistent with a lower tumor burden in patients prior to the onset of resistance. However, at progression this was reversed, with 21/30 (70%) samples demonstrating ctDNA evidence by aCGH or AR sequencing, including 9/30 (30%) with AR copy number gain and 13/30 (43%) with AR mutation (Figure 3A, eTable 2 in Supplement 2). Compared to baseline samples, we detected emergence and/or regression of particular copy number changes and AR mutations in over half the cohort (Figure 3A), suggesting tumor clonal population changes and consistent with complex and dynamic intra-patient heterogeneity^22–24. While this could be explained in some cases by absence of ctDNA in one time point, the majority of such patients had ctDNA evident at both baseline and progression. Furthermore, the median allelic frequency of AR mutations at baseline and progression was comparable (Figure 1A; eFigure 6 in Supplement 1), implying that the average ctDNA burden was not significantly higher at progression than baseline.

Figure 3.

Clinically-informative genomic aberration detectable in the cfDNA of mCRPC patients at progression on enzalutamide. a) Integrative landscape of alterations from aCGH, AR gene sequencing, and targeted sequencing. Columns represent patient samples at progression on enzalutamide, ordered in brevity of progression-free survival on treatment. *no baseline sample available for VC-093. b) Cluster of activating mutations falling within the phosphorylation domain of CTNNB1. Bar plot below indicates that this is a pan-cancer hotspot region for mutations. †two mutations detected in the same patient (VC-017). c) Germline defects in DNA damage repair genes detected across 30 patients at progression on enzalutamide. Protein schematic shows locations of BRCA2 mutations. Bar plot below demonstrates evidence of somatic loss-of-heterozygosity in patient cfDNA at progression.

Consistent with the original description of enzalutamide activity, we saw complete regression of clones bearing the bicalutamide-associated W742L/C mutations²⁵. Indeed, in an independent case we detected emergence of W742L/C in cfDNA during bicalutamide treatment, that dramatically regressed at therapy switch to enzalutamide and abiraterone (eFigure 7 in Supplement 1). Other mutations did not consistently regress; in fact, in patients with the same mutation detected at baseline and progression, 7/9 had an increase in detection frequency. Interestingly, we observed emergence of L702H positive clones in five patients (Figure 3A), and an increase in L702H frequency in three patients that were positive at baseline. Additionally, we did not see L702H regress in any patients with ctDNA evidence at progression. L702H has been linked to abiraterone resistance^13,15, but should be less relevant in the context of enzalutamide. Interestingly, all 11 patients with L702H at either baseline or progression had received prior abiraterone (11/40 vs. 0/22 p=0.0274; Fishers exact test). Similarly, T878A was only detected in patients that had received prior abiraterone (n=6) (6/40 vs. 0/22 p=0.08; Fisher’s exact test). In contrast, detection of AR copy number alterations was not associated with prior abiraterone. We detected only one instance of F877L (converts enzalutamide from antagonist to agonist²⁶) at progression, although the longest duration of response in patients with progression samples was 7 months.

Four patients with ≥2 AR mutations at baseline had a paired progression sample. Remarkably, three of these patients demonstrated regression or emergence of at least one (but never all) AR mutation(s) at progression, further suggesting intra-patient clonal heterogeneity. These patients progressed rapidly on enzalutamide with PFS of 1.7, 2.3 and 2.9 months respectively.

Clinically-actionable aberration in the cfDNA of progressing patients

A recent landmark study of mCRPC tissue biopsies suggests a large proportion of patients harbor specific genomic aberrations that potentially confer sensitivity to novel investigative agents¹⁸. We performed deep targeted sequencing of 19 frequently mutated or clinically-actionable prostate cancer genes in our progression samples with sufficient remaining cfDNA (n=14). Somatic mutations and/or copy number changes were detected in all samples, at a remarkably high allelic frequency (median mutation frequency excluding AR = 19.7%) (Figure 3A; eTable 3 and 4 in Supplement 2). Mutations were consistent with prostate cancer, including TP53 mutations (VC-019, VC-085, VC-093), a deleterious PTEN mutation (VC-057), and a missense mutation within the forkhead domain of FOXA1 (VC-056). Typical copy number changes, including deep deletions of PTEN (n=4/13), were robustly detected, with logratios that approach that expected of tissue biopsies (eTable 4 in Supplement 2). One patient (VC-085) with deep deletion of PTEN also harbored focal amplification of another PI3K pathway gene PIK3CB (eFigure 6 in Supplement 1), which may confer sensitivity to PIK3CB-specific inhibitors²⁷. Another patient (VC-019) with PTEN deletion had an activating p.E545K mutation in PIK3CA. We observed strong cross-platform validation for AR mutation, with all except one (low frequency) mutation redetected (eFigure 8 in Supplement 1).

We identified four patients with activating mutations in the serine phosphorylation domain of CTNNB1 (Figure 3B), including a patient with two different mutations at the same amino acid, suggesting a strong selective pressure. CTNNB1 mutations have been reported in primary prostate cancer (1%; TCGA²⁸) and front-line CRPC (3%¹⁸), but the elevated frequency in patients progressing rapidly on enzalutamide warrants follow-up and further clinical development of WNT pathway inhibitors²⁹.

Given the high frequency of germline DNA repair defects recently observed in advanced prostate cancer, we screened all patients at progression for germline BRCA1 and BRCA2 mutations (supplemented with additional DNA repair genes in four patients). Two patients harboured germline BRCA2 frameshift/nonsense mutations, and a third patient had a germline PALB2 frameshift mutation (Figure 3C; eFigure 9 in Supplement 1; eTable 5 in Supplement 2). One patient (VC-022) with a BRCA2 mutation had matched cfDNA sequenced, and loss-of-heterozygosity was evident in cfDNA (Figure 3C). These patients would be candidates for PARP inhibition or platinum chemotherapy^30,31, and our data suggests that cfDNA could serve as a powerful biomarker to identify bi-allelic BRCA2 loss.

Discussion

Liquid biopsies are a promising tool to develop predictive molecular biomarkers, but questions remain over applicability, since CTC studies require purification equipment, and study of cfDNA can necessitate sensitive assays such as digital PCR. However, here we show that simple techniques including aCGH and single gene NGS, can robustly detect ctDNA in >75% of mCRPC patients treated with enzalutamide. Furthermore, the average allelic frequency of AR mutation at baseline was >5%, suggesting that most second-line mCRPC patients have a ctDNA burden well above detection threshold for NGS approaches. Importantly, these results passed rigorous intra- and inter-platform validation.

AR copy number gain can drive CRPC progression, and its detection in cfDNA is linked to abiraterone progression^13,15. Although our aCGH analysis was unlikely to capture all samples with AR copy gain, we observed an association between AR amplification detection and poor outcomes on enzalutamide. This suggests that elevated AR expression levels (from gene amplification) are sufficient to overcome the potent inhibition of enzalutamide²⁵. It is also likely that elevated gene transcription (or rearrangements accrued during amplification) results in increased generation of constitutively-active truncated AR variants^32,33. This raises the possibility that in addition to AR amplification, patients with rapidly progressive disease on enzalutamide may have AR variants identified through CTC analysis^12,34.

Most patients with AR mutations did not exhibit primary resistance to enzalutamide. However, the persistence and emergence of clones harboring L702H and T878A mutations at progression suggests their continuing fitness during enzalutamide administration. This is surprising as AR-L702H has lower androgen affinity than wildtype AR³⁵, and both L702H and T878A are associated primarily with abiraterone resistance due to agonism by glucocorticoids and progesterones respectively^13,15,36. Since abiraterone-progressing patients commencing enzalutamide will frequently continue to receive prednisone (at least temporarily), the most parsimonious explanation is that the persistence of prednisone in circulation, together with endogenous adrenal androgens, continues to stimulate L702H- and T878A-positive clones. The implication that prednisone may contribute to both abiraterone and enzalutamide resistance means that a dexamethasone switch may prove efficacious³⁷, and that clinical development of new synthetic oral glucocorticoids (especially those without the 17α-OH group that increases affinity to L702H³⁵) is warranted^38,39.

Tumor heterogeneity is increasingly recognized as confounding factor for treatment of prostate cancer^22–24. We observed multiple AR mutations in ~8% of patients, and temporal AR mutation heterogeneity in several more, possibly reflecting tumor clones with distinct evolutionary niches. Furthermore, patients with multiple AR mutations at baseline performed poorly on enzalutamide. Given the demonstrated feasibility of AR mutation detection in cfDNA, analogues of our AR NSG assay, or digital PCR assays to detect specific mutations, could be easily implemented in future clinical trials (e.g. to prioritize patients for AR BF3 or DBD inhibitors^40,41). MYC gain, MET gain and RB1 loss were also associated with adverse outcomes. Although detection of these aberrations may simply reflect higher tumor burden in some patients, or indicate general genomic instability⁴², the association of RB1 loss with poor outcomes on enzalutamide is particularly interesting in light of pre-clinical data implicating RB1 loss in sensitivity to taxane chemotherapy⁴³.

We detected high allelic frequency mutations in CTNNB1 in 4/13 patients at progression. They mutate serine phosphorylation sites, preventing ubiquitination and degradation of CTNNB1. CTNNB1 mutations leading to WNT pathway activation are well-documented in several cancers, but are relatively rare in primary prostate cancer and even first-line CRPC (1% and 3% respectively)^18,28. However, WNT signaling is increasingly linked to therapy resistance in CRPC⁴⁴, and it is conceivable that CTNNB1 mutations confer increased cancer cell plasticity in the context of enzalutamide treatment. If validated as a resistance mechanism, cfDNA profiling could allow prospective identification of CRPC patients most likely to respond to porcupine inhibition²⁹. We also observed a high frequency of germline DNA repair defects in patients that progressed rapidly on enzalutamide, with somatic LOH detectable in cfDNA. The readiness of detection in cfDNA suggests a simple and practical method for prioritization of patients for PARP inhibition or DNA damaging agents³⁰.

Limitations of this study include relatively small sample size and inability to detect AR splice variants using cfDNA. In future it will also be important to examine resistance mechanisms in patients enjoying durable responses. 20/65 patients in our study remain on treatment, and are the most likely to develop enzalutamide-specific resistance mechanisms, including the F877L mutation that confers agonist activity to enzalutamide^26,45. Although multi-gene sequencing was not possible in all our samples (due to insufficient cfDNA after aCGH), our study demonstrates a series of robust assays that could be easily implemented in larger multi-centre cohorts to validate our findings and further develop cfDNA as a biomarker for improved clinical management of mCRPC.

Overall, this study demonstrates that clinically-informative genomic profiling from a minimally-invasive blood draw is feasible in nearly all patients with mCRPC. Beyond the molecular landscape and clinical associations reported here in the context of enzalutamide treatment, cfDNA therefore holds remarkable promise for the practical implementation of precision medicine programs in advanced prostate cancer.