Elucidating Prostate Cancer Behaviour During Treatment via Low-pass Whole-genome Sequencing of Circulating Tumour DNA

Abstract

Background

Better blood tests to elucidate the behaviour of metastatic castration-resistant prostate cancer (mCRPC) are urgently needed to drive therapeutic decisions. Plasma cell-free DNA (cfDNA) comprises normal and circulating tumour DNA (ctDNA). Low-pass whole-genome sequencing (lpWGS) of ctDNA can provide information on mCRPC behaviour.

Objective

To validate and clinically qualify plasma lpWGS for mCRPC.

Design, setting, and participants

Plasma lpWGS data were obtained for mCRPC patients consenting to optional substudies of two prospective phase 3 trials (FIRSTANA and PROSELICA). In FIRSTANA, chemotherapy-naïve patients were randomised to treatment with docetaxel (75 mg/m²) or cabazitaxel (20 or 25 mg/m²). In PROSELICA, patients previously treated with docetaxel were randomised to 20 or 25 mg/m² cabazitaxel. lpWGS data were generated from 540 samples from 188 mCRPC patients acquired at four different time points (screening, cycle 1, cycle 4, and end of study).

Outcome measurements and statistical analysis

lpWGS data for ctDNA were evaluated for prognostic, response, and tumour genomic measures. Associations with response and survival data were determined for tumour fraction. Genomic biomarkers including large-scale transition (LST) scores were explored in the context of prior treatments.

Results and limitations

Plasma tumour fraction was prognostic for overall survival in univariable and stratified multivariable analyses (hazard ratio 1.75, 95% confidence interval 1.08–2.85; p = 0.024) and offered added value compared to existing biomarkers (C index 0.722 vs 0.709; p = 0.021). Longitudinal changes were associated with drug response. PROSELICA samples were enriched for LSTs (p = 0.029) indicating genomic instability, and this enrichment was associated with prior abiraterone and enzalutamide treatment but not taxane or radiation therapy. Higher LSTs were correlated with losses of RB1/RNASEH2B, independent of BRCA2 loss.

Conclusions

Plasma lpWGS of ctDNA describes CRPC behaviour, providing prognostic and response data of clinical relevance. The added prognostic value of the ctDNA fraction over established biomarkers should be studied further.

Patient summary

We studied tumour DNA in blood samples from patients with prostate cancer. We found that levels of tumour DNA in blood were indicative of disease prognosis, and that changes after treatment could be detected. We also observed a “genetic scar” in the results that was associated with certain previous treatments. This test allows an assessment of tumour activity that can complement existing tests, offer insights into drug response, and detect clinically relevant genetic changes.

Take Home Message

Tumour fraction, independent from cell-free DNA concentration and other clinical variables, is prognostic, with low-pass whole-genome sequencing profiles detecting abiraterone/enzalutamide-induced emerging genomic instability that is likely to be acquired as a result of tumour development.

1. Introduction

Prostate cancer (PC) remains a major cause of male cancer deaths [1] despite significant advances in systemic treatment [2]. Taxanes improve overall survival (OS) and provide symptomatic benefit from metastatic castration-resistant PC (mCRPC) [3], [4] but uncertainty remains regarding how long to continue treatment, with early prostate-specific antigen (PSA) changes unable to guide treatment switch decisions. Better biomarkers of response are needed for early discontinuation of ineffective treatment. Measurement of total circulating cell-free DNA (cfDNA) in serial plasma samples is prognostic but is challenged by nonmalignant circulating DNA. The fraction of cfDNA that is tumour-derived—circulating tumour DNA (ctDNA)—can be estimated via next-generation sequencing to generate quantitative [5] and qualitative data [6], with serial ctDNA genomics providing important insights into disease behaviour and evolution [7].

Low-pass whole-genome sequencing (lpWGS) assesses genome-wide copy number events [8], allowing rapid, high-throughput, inexpensive testing of ctDNA and estimation of cfDNA tumour fraction, ploidy, and whole-genome copy number alterations (CNAs). lpWGS can characterise the impact of DNA repair defects and genomic instability by detailing copy-number fragment size and frequency, including large-scale transitions (LSTs) and tandem duplications [9], [10], [11]. Homologous repair scores, to which LSTs contribute, have been linked to PARP inhibitor sensitivity, and tandem duplication CNA genotypes reflect deleterious CDK12 alterations that sensitise the individual to immunotherapy [10], [12].

We validated a cfDNA lpWGS assay and tested samples from mCRPC patients treated in two taxane phase 3 trials, FIRSTANA (NCT01308567) and PROSELICA (NCT01308580). In FIRSTANA, chemotherapy-naïve patients were randomised to docetaxel (75 mg/m²) or cabazitaxel (20 or 25 mg/m²); in PROSELICA, docetaxel-pretreated patients were randomised to 20 or 25 mg/m² cabazitaxel [13], [14]. Here we present lpWGS data from patients consenting to these extra analyses and clinical qualification that this biomarker provides prognostic and response data as well as information on evolving progressing disease.

2. Patients and methods

2.1. Patients and sample collection

Supplementary Figure 1 shows an outline of our study, while Supplementary Figure 2A,B shows overviews of the FIRSTANA and PROSELICA trials. Both trials have already been reported [13], [14], [15]. OS was defined as the time from randomisation to death from any cause. Patients who did not have an event were censored at the date of last contact. Radiographic and PSA progression-free survival (rPFS and PSA-PFS) were defined as the time from randomisation to the date of tumour progression.

To determine progression and response status we used Response Evaluation Criteria in Solid Tumours 1.1 and Prostate Cancer Working Group 2 Criteria, as previously described in the FIRSTANA and PROSELICA reports [13], [14], [15]. Studies on ctDNA were prospectively included as an optional exploratory endpoint in both trials. Clinical data included PSA, lactate dehydrogenase, haemoglobin, serum albumin, alkaline phosphatase, and Eastern Cooperative Oncology Group performance status. Plasma was collected from patients consenting to this substudy at two baseline time points 1–4 wk apart (at screening [SCR] and cycle 1 day 1 [C1]) and at cycle 4 day 1 (C4) and at the end of the study (EOS). Blood was collected in lithium heparin tubes (BD Vacutainer; BD Biosciences, San Jose, CA, USA). Plasma was collected from healthy volunteers (n = 10) and pooled prior to lpWGS. Biopsies were collected according to a prospective PC molecular characterisation protocol approved by the institutional review board [16].

2.2. cfDNA extraction and quantification

cfDNA was isolated from 1–4 ml of plasma using a QIAamp Circulating Nucleic Acid kit (Qiagen, Hilden, Germany). Of the 60 μl of eluate, 3 μl was used for quantification via a Quant-IT Picogreen HS DNA kit (ThermoFisher, Waltham, MA, USA) and concentrations were read on a BioTek microplate spectrophotometer (excitation 480 nm, emission 520 nm) (BioTek Instruments, Winooski, VT, USA).

2.3. Library preparation and sequencing

Following extraction, samples were treated with heparinase I (Sigma-Aldrich, St. Louis, MO, USA) [17]. Low-pass whole-genome library preparation was carried out using a Qiagen QiaSeq FX DNA library kit (Qiagen). Samples were sequenced on an Illumina NovaSeq 6000 platform (Illumina, San Diego, CA, USA). Technical replicates were prepared and sequenced on an Illumina MiSeq system (Illumina) [18].

2.4. Bioinformatics processing

lpWGS data were converted to paired-end reads (bcl2fastq2 v.2.17.1.14) with default settings and subsequently aligned to the human reference genome (GRCh37) using the BWA-MEM (version 0.7.12) algorithm [33]. Quality control checks were performed using Picard (Broad Institute, Cambridge, MA, USA, version 2.8.1) and FASTQC (Babraham Institute, Babraham, UK, version 0.11.8). Samples were excluded from analysis if the sequencing depth was less than 0.05× or if they failed the FASTQC read quality filter. Aligned reads were quantified using HMMcopy readCounter [35] (version 0.99.0) with the quality filter and interval width set to 20 and 500 kb, respectively.

Read depth data were modelled and the tumour fraction was calculated using ichorCNA (version 0.1.0) [19]. Transition strength parameters were set at --txnE = 0.99999 and --txnStrength = 100000; the maximum copy number (CN) was set to 20 to account for amplifications. The germline DNA fraction (initial values 40% and 90%), ploidy (initial values 2 and 3), and subclonality were modelled. The default 500-kb reference coverage data set supplied with ichorCNA was used.

2.5. Data handling and statistical analyses

IchorCNA provides segmented CN data and estimates of tumour fraction and ploidy. These were further processed to generate both CN calls and LST values, as described in the Supplementary material [20]. Cox proportional-hazards models were used for multivariable survival analysis, and Kaplan-Meier curves for univariable survival comparisons. The log-likelihood test was used to compare regression models. Multivariable generalised linear and logistic regressions were used to study continuous and binary outcomes, respectively. A full listing of the regression variables is provided in the Supplementary material. Comparisons of continuous variables (tumour fraction and LST scores) between groups (study subset, prior treatment, specific CN events) in violin plots used the Wilcoxon rank-sum (unpaired data) or signed-rank (paired data) tests. Elastic-net regression was performed using eNetXplorer (version 1.1.0) [21], for which parameter selections are presented in the Supplementary material. Following elastic-net regression, significant (p < 0.05) bins were merged with adjacent highly correlated (Pearson’s r > 0.9) bins to form merged regions for subsequent analyses, with the CN values of the most significant overlapping bin assigned to each final merged region.

3. Results

3.1. FIRSTANA and PROSELICA cohorts providing plasma samples for lpWGS

lpWGS data were generated from 540 samples acquired at four time points (SCR and C1, representing baseline, and C4 and EOS). Samples with plasma available after other preplanned analyses were tested, including 299 FIRSTANA (104 SCR, 55 C1, 79 C4, and 61 EOS) and 241 PROSELICA (84 SCR, 34 C1, 59 C4, and 64 EOS) samples. The characteristics for the 188 patients involved are presented in Table 1 (FIRSTANA n = 103 and PROSELICA n = 85; Supplementary Figs. 1 and 2). PROSELICA substudy patients had worse prognostic characteristics than the FIRSTANA group, in keeping with a more heavily pretreated cohort. Only <2% FIRSTANA substudy patients received abiraterone/enzalutamide, in contrast to 34% of PROSELICA substudy patients. The PSA response rate was lower in PROSELICA than in FIRSTANA (39% vs 66%), similar to previous reports [5], [13], [14]. The baseline clinical and survival characteristics of these cohorts compared to the overall trial populations are described in Supplementary Table 1. Maximum follow-up across the two trials was 51 and 48 mo. No differences in response or survival were noted in the substudy cohorts compared to the overall populations.

Table 1.

Baseline characteristics of the cohorts with castration-resistant prostate cancer from the FIRSTANA and PROSELICA studies

Characteristic	FIRSTANA (n = 103)	PROSELICA (n = 85)	p value
ECOG PS ≥2, n (%) b	3 (2.9)	10 (12)	0.02 a
RECIST-measurable, n (%) b	54 (52)	45 (53)	0.9 a
Visceral disease, n (%)	20 (19)	25 (29)	0.11 a
Pain at baseline, n (%) c	69 (67)	59 (69)	0.4 a
Gleason score <8 at diagnosis, n (%) d	42 (41)	45 (53)	0.051 a
Prior Abi/Enza treatment, n (%)	2 (1.9)	29 (35)	<0.001 a
Trial arm, n (%)
Cabazitaxel (20 mg/m2)	34 (33)	41 (48)
Cabazitaxel (25 mg/m2)	28 (27)	44 (52)
Docetaxel (75 mg/m2)	41 (40)	0 (0)
Median age, yr (IQR)	68.0 (62.5–72.0)	67 (64.0–71.0)	0.2 e
Median LDH, U/l (IQR)	267 (204–360)	366 (234–605)	0.003 e
Median ALP, U/l (IQR)	129 (81.0–242)	214 (118–413)	0.002 e
Median haemoglobin, g/dl (IQR)	121 (111–128)	112 (105–122)	<0.001 e
Median albumin, g/dl (IQR)	40.2 (37.5–43.4)	40.0 (36.0–43.0)	0.2 e
Median PSA, ng/ml (IQR)	60.4 (18.2–188)	161 (64.9–623)	<0.001 e
Median PSA doubling time, d (IQR)	62 (36–100)	51 (35–86)	0.3 e
Median NLR (IQR)	2.25 (1.53–4.10)	2.79 (1.69–3.89)	0.17 e
Outcomes
>50% PSA response at 12 wk, n (%)	55 (53)	23 (27)	<0.001 a
>50% PSA response at any time, n (%)	68 (66)	33 (39)	<0.001 a
Median OS, mo (95% CI)	21.3 (17.2–23.9)	13.3 (11.5–15.8)	<0.001 f
Median rPFS, mo (95% CI)	11.6 (9.86–13.6)	7.13 (6.11–12.4)	0.003 f
Median PSA-PFS, mo (95% CI)	7.43 (6.64–8.94)	4.86 (4.14–7.29)	0.003 f
Censored (n)	30	6
Median FU for censored patients (mo)	32.0	Insufficient cases

Abi = abiraterone; ALP = alkaline phosphatase; ECOG PS = Eastern Cooperative Oncology Group performance status; Enza = enalutamide; FU = follow-up; IQR = interquartile range; LDH = lactate dehydrogenase; NLR = neutrophil/lymphocyte ratio; OS = overall survival; PFS = progression-free survival; PSA = prostate-specific antigen; RECIST = Response Evaluation Criteria in Solid Tumours; rPFS = radiographic PFS.

^aχ² test.

^bStratification parameters.

^cTwenty assessments missing (15 in FIRSTANA and 5 in PROSELICA).

^dTwelve assessments missing (5 in FIRSTANA and 7 in PROSELICA).

^eWilcoxon rank-sum test.

^fLog-rank test.

3.2. Validation of lpWGS CNA profile acquisition

lpWGS data were generated from cfDNA samples, with median sequencing coverage of 1.9X . We identified mCRPC genomic profiles (Supplementary Fig. 3) with common events including AR (chromosome Xq) and MYC (chromosome 8q) loci (Supplementary Fig. 3A). Cases with frequent large-scale CN changes described in tumours with BRCA1/2 loss were observed [22] (Supplementary Fig. 3B), as well as focal tandem duplication patterns linked to CDK12 loss [10] (Supplementary Fig. 3C). The overall CNA profile of our cohort was consistent with previously published mCRPC biopsy data [23] (Supplementary Fig. 3D).

To evaluate lpWGS CNA sensitivity, we used a baseline sample with an estimated tumour fraction of 50% and serially diluted it with pooled healthy volunteer cfDNA. The diluted profiles closely matched each other, but estimation of absolute CNA values was challenging when the tumour fraction was <5% (Supplementary Fig. 4A) [19]. cfDNA tumour fractions estimated via lpWGS were consistent with the expected fractions following dilution (Pearson’s r = 0.994; Supplementary Fig. 4B). We further studied technical and biological replicates; ten samples subjected to duplicate library preparations showed highly correlated CNA profiles (Pearson’s r = 0.948; Supplementary Fig. 4C). Biological replicates (same-patient samples taken 1–4 wk apart at SCR and C1, n = 88) showed reproducible CNAs (Supplementary Fig. 5A) and tumour fractions (Supplementary Fig. 5B).

3.3. Baseline cfDNA tumour fraction and prognosis

To explore univariable associations, we split the entire FIRSTANA + PROSELICA lpWGS cohort according to the overall median baseline ctDNA fraction, which was 9.6% (Fig. 1A). High baseline ctDNA fraction was correlated with shorter OS in the FIRSTANA and PROSELICA trials, with 10-mo-shorter median OS (Fig. 1A,B). Analysis of the tumour fraction as a continuous variable in FIRSTANA and PROSELICA independently and collectively confirmed these findings (Cox proportional-hazard models, p < 0.001 and p = 0.004; meta-analysis test for overall effect, p < 0.001; Fig. 1C). The baseline ctDNA fraction was also significantly associated with rPFS and PSA-PFS (Supplementary Fig. 6). The relationship between detectable ctDNA fraction (present/absent) on treatment and OS is shown in Supplementary Figure 7. Furthermore, the baseline ctDNA fraction was associated with several other clinical variables (Supplementary Fig. 8).

Fig. 1.

Cell-free DNA (cfDNA) tumour fraction estimates are prognostic. (A,B) Univariable analysis of overall survival (OS) with cohort stratification by median baseline tumour fraction (average for the screening and cycle 1 day 1 samples; yellow = high, blue = low) across the (A) FIRSTANA (discovery) and (B) PROSELICA (validation) cohorts. Kaplan-Meier plots with confidence intervals (CIs) and matching risk tables are shown. Dashed lines indicate the time to 50% survival. (C) Univariable regression random-effects meta-analysis of OS by cfDNA tumour fraction (continuous variable, log₁₀ transformed) in the FIRSTANA (discovery) and PROSELICA (validation) cohorts. Effect size, hazard ratio (HR), 95% CI, and p values for the overall test for effect are shown. SE = standard error. (D) Forest plots showing multivariable analysis of OS, radiographic progression-free survival (RPFS) and PSA progression-free survival (PSAPFS) from Cox proportional-hazards models. Statistical models were stratified by study inclusion (FIRSTANA vs PROSELICA) because of underlying differences in survival. The median baseline tumour fraction (continuous variable, log₁₀-transformed) and median baseline cfDNA total concentration (continuous variable, log10-transformed) are included in the model, along with other baseline clinical variables prognostic for mCRPC.

ECOG = Eastern Cooperative Oncology Group; Mets = metastases; ALB = albumin; HB = haemoglobin; LDH = lactate dehydrogenase; PSA = prostate-specific antigen; ALP = alkaline phosphatase; NLR = neutrophil/lymphocyte ratio.

We applied a multivariable Cox proportional-hazards model, as described in the Supplementary material, to investigate OS, rPFS, and PSA-PFS (n = 184 of 188 patients with complete data) and, importantly, stratified by study inclusion (Fig. 1D) [5]. The baseline ctDNA fraction was independently associated with worse outcomes for all three endpoints; the hazard ratio was 1.75 (95% confidence interval [CI] 1.08–2.85; p = 0.024) for OS, 3.29 (95% CI 1.65–6.59; p < 0.001) for rPFS, and 1.62 (95% CI 0.98–2.68; p = 0.061) for PSA-PFS. We found that inclusion of the tumour fraction in multivariable survival models led to a significant improvement (OS, C index 0.722 vs 0.709; likelihood-ratio test, p = 0.021; Supplementary Table 2). These data indicate that the ctDNA fraction provides independent information in the context of established clinical biomarkers.

3.4. ctDNA fraction and response to taxane treatment

There was no statistically significant difference in baseline ctDNA fraction between the FIRSTANA and PROSELICA cohorts (median 9.5% vs 10%; Wilcoxon rank-sum test, p = 0.2; Fig. 2A). The ctDNA fraction among nonresponding FIRSTANA patients (determined via serial radiological and PSA analyses) was relatively stable throughout treatment (median 11% at baseline, 7.5% on treatment, and 10% at EOS). By contrast, the ctDNA fraction among responders exhibited large decreases on treatment (median 8.5% at baseline, 0% on treatment, and 3.7% at EOS). These observations were replicated in the PROSELICA cohort (Fig. 2B). In keeping with these findings, CNAs became undetectable in responders while on treatment (Supplementary Fig. 9). We further evaluated the ctDNA fraction in longitudinal same-patient samples (n = 252 samples from 84 patients; Fig. 2C). Among responders, the ctDNA fraction significantly decreased on treatment (Wilcoxon signed-rank test, p < 0.001) and increased at EOS (median 8.7% at baseline, 0% on treatment, and 7.2% at EOS); this was not observed for nonresponders (median 9.7% at baseline, 9.1% on treatment, and 11.3% at EOS).

Fig. 2.

Treatment-induced changes in the tumour fraction in cell-free DNA (cfDNA). (A) Violin plots comparing baseline tumour fraction between FIRSTANA (FIRS) and PROSELICA (PROS) plasma samples. Wilcoxon rank-sum test: p = 0.2. Values for the median (white point; FIRS = 0.0949, PROS = 0.101), interquartile range (IQR, black rectangle; FIRS = 0.0436 to 0.175, PROS = 0.0740 to 0.197), and 1.5 × IQR (black lines) tumour fractions are also shown. (B) Violin plots illustrating changes in tumour fraction in the unmatched cohort for patients categorised as responders (either prostate-specific antigen decrease of 50% or radiographic response) or nonresponders at multiple time points. Plots are split between FIRS and PROS cohorts. Unpaired Wilcoxon rank-sum tests yielded the following results: baseline (BL) FIRS: responder (n = 70, median 0.0850) versus nonresponder (n = 33, median 0.112), p = 0.045; on treatment (OT) FIRS: responder (n = 55, median 0) versus nonresponder (n = 22, median 0.0755), p = 0.003; end of study (EOS) FIRS: responder (n = 39, median 0.0365) versus nonresponder (n = 18, median 0.105), p = 0.035; BL PROS: responder (n = 34, median 0.0976) versus nonresponder (n = 51, median 0.109), p = 0.13; OT PROS: responder (n = 26, median 0.0213) versus nonresponder (n = 32, median 0.0991), p < 0.001; EOS PROS: responder (n = 22, median 0.0766) versus nonresponder (n = 41, median 0.134), p = 0.035. (C) Analysis of matched same-patient cfDNA sample sets showing longitudinal tumour fraction changes for nonresponders (n = 39) and responders (n = 45). Median values are in black. Lines connecting points indicate same-patient sets. Paired Wilcoxon signed-rank tests yielded the following results: nonresponder BL (median 0.0966) versus OT (median 0.091), p = 0.06); nonresponder BL (median 0.0966) versus EOS (median 0.113), p = 0.17; nonresponder OT (median 0.0912) versus EOS (median 0.113), p = 0.006; responder BL (median 0.0873) versus OT (median 0), p < 0.001; responder BL (median 0.0873) versus EOS (median 0.0718), p = 0.2; responder OT (median 0) versus EOS (median 0.0718), p = 0.001.

3.5. lpWGS provides insights into PC genomic instability

We next explored LST as a measure of genomic instability [22]. PROSELICA baseline samples exhibited a significantly higher LST score compared with FIRSTANA baseline samples (Wilcoxon rank-sum test, p = 0.029; Fig. 3A); however, there was no correlation between LST score and estimated tumour fraction (Supplementary Fig. 10A). We hypothesised that LST score differences could be explained by prior treatment, and found that patients who had received abiraterone or enzalutamide had significantly higher LST scores (median LST score: 18 untreated vs 26 treated; Wilcoxon rank-sum test, p = 0.003; Fig. 3B and Supplementary Fig. 10B). Examples of CN profiles of samples with high, intermediate, and low LST scores are shown in Supplementary Figure 10C.

Fig. 3.

Genomic copy number burden and large-scale transition (LST) in metastatic castration-resistant prostate cancer (mCRPC). (A) Violin plot comparison of LST scores between baseline (BL) and end of study (EOS) for the FIRSTANA (FIRS) and PROSELICA (PROS) cohorts. Samples with a tumour fraction >5% were assessed; in cases with multiple BL data, the highest-fraction sample was used. Values for the median (white point), interquartile range (IQR; black rectangle) and 1.5 × IQR (black lines) LST scores are also shown. Unpaired Wilcoxon rank-sum tests yielded the following results: FIRS BL (median 18, IQR 11-24.5) versus FIRS EOS (median 18, IQR 9-29), p = 0.5; FIRS BL (median 18, IQR 11-24.5) versus PROS BL (median 20, IQR 14-31), p = 0.029; PROS BL (median 20, IQR 14-31) versus PROS EOS (median 22, IQR 15-33), p = 0.7. (B) Comparison of baseline LST scores between patients with prior abiraterone (Abi) or enzalutamide (Enza) exposure and untreated patients (n = 152). FIRSTANA patients are shown in green and PROSELICA patients in purple; after removal of cases with a low tumour fraction, no FIRSTANA patients had received Abi/Enza and a significant proportion of PROSELICA patients had received Abi/Enza. Median, IQR, and 1.5 × IQR LST scores shown as for A. Unpaired Wilcoxon rank-sum tests yielded the following results: untreated FIRS (median 18, IQR 11-24.5) versus untreated PROS (median 18, IQR 13-27.5), p = 0.3; overall untreated (median 18, IQR 11.2-25) versus overall treated (median 26, IQR 15.2-36.8), p = 0.003. (C) Forest plot of the multivariable logistic regression model for association of LST with prior Abi/Enza treatment, depicting ability of LST score to predict Abi/Enza status in the context of tumour fraction and other clinical biomarkers. (D) Comparison of baseline LST values between patients who had received prior radical radiotherapy and those who had not. FIRSTANA patients are shown in green and PROSELICA patients in purple (n = 152). Median, IQR, and 1.5 × IQR LST scores shown as for A. Unpaired Wilcoxon rank-sum tests yielded the following results: untreated FIRS (median 18, IQR 12-25) versus untreated PROS (median 19.5, IQR 14-31), p = 0.10; treated FIRS (median 17, IQR 8.25-24) versus treated PROS (median 20, IQR 14-35), p = 0.11; overall untreated (median 18, IQR 12-26) versus overall treated (median 19, IQR 12.5-27.5), p = 0.8. (E) Violin plot comparison of baseline LST scores in FIRSTANA and PROSELICA by taxane treatment response (n = 152). Median, IQR, and 1.5 × IQR LST scores shown as for A. A Wilcoxon rank-sum test for nonresponders (median 18, IQR 12-28) versus responders (median 19, IQR 12-25) yielded p = 0.6. (F) Comparison of LST scores in the validation cohort of same-patient samples (44 patients, 88 samples) between primary hormone-sensitive prostate cancer (HSPC) and mCRPC sample pairs. Median, IQR, and 1.5 × IQR LST scores shown as for A. A paired Wilcoxon rank-sum test for HSPC (median 18, IQR 12-22) versus mCRPC (median 22.5, IQR 17.8-30.8) yielded p < 0.001.

ECOG = Eastern Cooperative Oncology Group; Mets = metastases; OR = odds ratio; CI = confidence interval.

This was maintained in multivariable logistic regression analysis that included other known prognostic factors (including ctDNA tumour fraction) with the LST score (odds ratio 17.2, 95% CI 1.98–223; p = 0.018; Fig. 3C). However, there was no significant difference in LST score between individuals pretreated with radical radiotherapy and untreated individuals (Fig. 3D), or between taxane responders and nonresponders (Fig. 3E).

We confirmed these findings using same-patient samples from hormone-sensitive PC (diagnostic, primary tumours) and CRPC (metastatic) biopsies (n = 88) [16], all of whom had received abiraterone or enzalutamide before CRPC biopsy (Supplementary Table 3). LST scores increased significantly (median scores 18 and 22.5; Wilcoxon rank-sum test, p < 0.001) after exposure to abiraterone or enzalutamide in this cohort (Fig. 3F).

3.6. Association of LST score with tumour CNAs

To explore CNAs associated with LST scores, we performed elastic-net regression, which avoids false positives due to multiple testing. Among baseline samples, several CNAs were associated with high LST values (p < 0.05), the most significant of which was chromosome 13 loss (Fig. 4A). We then applied a multivariable generalised linear model to assess the association of these regions with LST score and found that loss of two regions on chromosomes 13 and 17, and gain of two regions on chromosomes 1 and 20 remained the most strongly associated with high LST scores (Fig. 4B).

Fig. 4.

Genomic regions associated with LST. (A) Elastic-net regression model results for genome-wide copy-number alterations showing cross-validated p values (−log₁₀-transformed) and regression coefficients for each 500-kb genomic bin. Loci with significant (p < 0.05) genomic alteration are depicted in red in the p-value plot and in green in the regression-coefficient plot. (B) Forest plot showing multivariable generalised linear model analysis of the genomic loci identified as associated with higher LST scores. Tumour fraction (log₁₀-transformed), prior abiraterone (Abi) or enzalutamide (Enza) treatment, and study inclusion (FIRSTANA vs PROSELICA) were also included in the model. (C) Heatmap illustrating CNAs in regions (rows) significantly (p < 0.05) associated with LST scores. The samples (columns) are ordered by LST score, with additional annotation for estimated tumour fraction, prior abiraterone/enzalutamide status, and study inclusion (FIRSTANA vs PROSELICA). CNAs are coloured by type.

LST = large-scale transition; CNA = copy number alteration; Del = deletion.

Genes within these loci, with an overlap of publicly available PARP inhibitor CRISPR screen data, are detailed in the Supplementary material. Interestingly, when ranked by LST score, co-occurrence of several of the genomic loci most associated with a high LST score was evident (Fig. 4C). However, we surprisingly observed that the key loci highlighted by this multivariable approach did not point to BRCA2, but instead focused on the RB1/RNASEH2B region, although these analyses are limited by sample size and require further validation (Fig. 5A,B). Overall, these data indicate that an increase in LST score and genomic instability emerge following treatment with abiraterone or enzalutamide, but not taxane, and this is associated with RB1 genomic locus loss with and without BRCA2 genomic loss, with possible implication of concurrent RNASEH2B genomic loss.

Fig. 5.

Chromosome 13 deletions and LST score. Violin plot comparisons of baseline LST scores for different chromosome 13 deletion combinations across the (A) FIRSTANA and (B) PROSELICA cohorts. Values for the median (white point), interquartile range (IQR; black rectangle), and 1.5 × IQR (black lines) LST scores are shown. The p values are for between-group Wilcoxon rank-sum tests.

lpWGS = low-pass whole-genome sequencing; LST = large-scale transition.

4. Discussion

Our data show that the ctDNA fraction obtained via lpWGS provides information regarding prognosis, with serial analyses identifying responding disease. Critically, these assays provide genomic information that can serve as biomarkers of treatment response and detect genomic aberrations and signatures of DNA repair defects [10]. ctDNA lpWGS provides information on emerging mCRPC genomics and can be performed frequently, yielding data for multiple metastatic sites. However, ctDNA assays do have limitations and cannot fully elucidate intrapatient disease heterogeneity; this may be more feasible with analysis of single circulating tumour cells [24].

Interestingly, our study also shows that treatment with abiraterone or enzalutamide is associated with increases in CNAs and LST scores; this was not seen following taxane treatment or radical radiotherapy, and was maintained in the presence of other markers of tumour burden. This suggests differential, therapy-induced, genomic alterations following treatment with next-generation hormonal agents, but not taxanes. Other studies have shown that high LST scores are associated with genomic instability and may be biomarkers for PARP inhibitor sensitivity [12]. We show here that the increase in LST score following treatment with abiraterone or enzalutamide is most associated with loss of loci on chromosome 13, which usually contains the RB1 and RNASEH2B genes, among others. Tumours with loss of RNASEH2B have been linked to impaired misincorporated RNA excision from DNA and can sensitise to PARP inhibition [25].

A limitation of ctDNA lpWGS is the fact that plasma samples containing low tumour fractions make precise detection of CNAs, especially subclonal events, challenging [19], [26]. A low ctDNA fraction is nevertheless of clinical significance and indicates lower tumour burden and better outcomes. The variation in ctDNA values observed over time may reflect disease behaviour [27]. The ctDNA fraction estimated via lpWGS, unlike cfDNA concentrations [14], is highly correlated with disease response and treatment outcome, probably because a significant proportion of cfDNA arises from nontumour sources [28], [29]. Our observations support the proposal that changes in the ctDNA fraction can track treatment response, and we welcome further exploration of this phenomenon [30], [31].

This study may be biased by sample selection; not all patients consented to blood sample donation for these analyses and our analyses focused on patients with available plasma with higher ctDNA levels, which may represent a poor-prognosis cohort. However, there were no statistically significant differences in the response rate or survival between our cohort and the overall trial populations. mCRPCs overall do have some of the highest quantities of cfDNA among adult solid tumour types, making these studies especially useful [32]. This study supports complementary research on ctDNA lpWGS and targeted next-generation sequencing, which together can transform disease management. Further studies to critically validate the prognostic capability of the ctDNA fraction in mCRPC are now necessary, especially in the context of existing validated biomarkers.

5. Conclusions

In conclusion, we demonstrated that cfDNA lpWGS can describe CRPC behaviour, provide prognostic and response data of clinical relevance, and identify emerging genomic alterations that might serve as therapeutic targets.



Author contributions: Johann S. de Bono had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.



Study concept and design: Sumanasuriya, Seed, Mehra, Eisenberger, Sartor, Oudard, Ozatilgan, Geffriaud-Ricouard, Chadjaa, Macé, de Bono.

Acquisition of data: Sumanasuriya, Christova, Pope, Bertan, Bianchini, Rescigno, Figueiredo, Goodall, Fowler, Flohr, Baxter, Pettitt.

Analysis and interpretation of data: Sumanasuriya, Seed, Pettit.

Drafting of the manuscript: Sumanasuriya, Seed, Rekowski, Lord, Yuan, de Bono.

Critical revision of the manuscript for important intellectual content: Sumanasuriya, Rescigno, Mehra, Neeb, Rekowski, Eisenberger, Sartor, Oudard, Ozatilgan, Geffriaud-Ricouard, Chadjaa, Macé, Lord, Lambros, Sharp, Mateo, Carreira, Yuan, de Bono.

Statistical analysis: Seed, Parr, Rekowski.

Obtaining funding: Eisenberger, Sartor, Oudard, Geffriaud-Ricouard, Chadjaa, Macé, de Bono.

Administrative, technical, or material support: None.

Supervision: Yuan, de Bono.

Other: None.



Financial disclosures: Johann S. de Bono certifies that all conflicts of interest, including specific financial interests and relationships and affiliations relevant to the subject matter or materials discussed in the manuscript (eg, employment/affiliation, grants or funding, consultancies, honoraria, stock ownership or options, expert testimony, royalties, or patents filed, received, or pending), are the following: Johann S. de Bono has served on advisory boards and received fees from many companies including AstraZeneca, Astellas, Bayer, Bioxcel Therapeutics, Boehringer Ingelheim, Cellcentric, Daiichi, Eisai, Genentech/Roche, Genmab, GSK, Janssen, Merck Serono, Merck Sharp & Dohme, Menarini/Silicon Biosystems, Orion, Pfizer, Qiagen, Sanofi Aventis, Sierra Oncology, Taiho, and Vertex Pharmaceuticals. He is an employee of the Institute of Cancer Research, which has received funding or other support for his research work from AstraZeneca, Astellas, Bayer, Cellcentric, Daiichi, Genentech, Genmab, GSK, Janssen, Merck Serono, MSD, Menarini/Silicon Biosystems, Orion, Sanofi Aventis, Sierra Oncology, Taiho, Pfizer, and Vertex, and which has a commercial interest in abiraterone, PARP inhibition in DNA repair defective cancers, and PI3K/AKT pathway inhibitors (no personal income). He was named as an inventor, with no financial interest, for patent 8,822,438 and has been the Chief Investigator/Principal Investigator of many industry-sponsored clinical trials. Chris Lord has received research funding from AstraZeneca, Merck KGaA, and Artios; has received consultancy, SAB membership, or honoraria payments from Syncona, Sun Pharma, Gerson Lehrman Group, Merck KGaA, Vertex, AstraZeneca, Tango, 3rd Rock, Ono Pharma, and Artios; has stock in Tango and Ovibio; and is a named inventor on patents describing the use of DNA repair inhibitors and stands to gain from their development as part of the Institute of Cancer Research “Rewards to Inventors” scheme. Joaquin Mateo has participated in advisory boards for Amgen, AstraZeneca, Clovis Oncology, Janssen, MSD, and Roche; has received research funding from AstraZeneca and Pfizer Oncology (not related to this work); and is the Principal Investigator for several industry-sponsored clinical trials. Niven Mehra has served on advisory boards (compensated and institutional) for Roche, MSD, BMS, Bayer, Astellas and Janssen; has received research support (institutional) from Astellas, Janssen, Pfizer, Roche, and Sanofi-Genzyme; and has received travel support from Astellas and MSD. Wei Yuan has received a meeting travel grant from Jilin Huarui Gene Technology Ltd. Christine Geffriaud-Ricouard, Ayse Ozatilgan, Mustapha Chadjaa, and Sandrine Macé are employees of Sanofi-Genzyme. The remaining authors have nothing to disclose.



Funding/Support and role of the sponsor: This work was supported by Prostate Cancer UK and the Movember Foundation through the London Movember Centre of Excellence (CEO13-2-002), Cancer Research UK (Centre Programme grant), Experimental Cancer Medicine Centre grant funding from Cancer Research UK and the Department of Health, and Biomedical Research Centre funding to the Royal Marsden ECMC (CRM064X). George Seed was funded by a Prostate Cancer UK PhD Studentship (grant ref. TLD-S15-006, 2016–2020) and a Prostate Cancer UK research fellowship (grant ref. TLD-PF19-005, from 2021). Semini Sumanasuriya was funded by a Prostate Cancer UK grant (ref. RIA1 5-ST2-O18). The authors are grateful for support and funding from Sanofi Aventis. The sponsor played a role in collection of the data and review and approval of the manuscript.



Acknowledgments: We thank all the patients and their families for participating in these studies and all the staff involved at the participating hospitals. Johann S. de Bono is a National Institute for Health Research (NIHR) Senior Investigator. The views expressed in this article are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health.

Associate Editor: Maarten Albersen

Appendix A. Supplementary data

The following are Supplementary data to this article: