Genomic features and response to PARP Inhibition in metastatic castration-resistant prostate cancer

Abstract

PURPOSE

PARP inhibitors (PARPis) are effective for the treatment of metastatic castration-resistant prostate cancer (mCRPC) harboring homologous recombination repair (HRR) gene alterations, but predictive biomarkers beyond individual gene alterations remain poorly defined. We aimed to identify genomic features associated with response to PARP inhibition in real-world data.

METHODS

This is a single-center retrospective study of patients with mCRPC who received PARPi therapy. All patients underwent tumor genomic profiling using a single panel sequencing assay. Responses were defined as PSA50 (≥50% PSA decline from baseline) and objective radiographic response by RECIST 1.1. We explored the association of genomic features with response to therapy.

RESULTS

Among 120 mCRPC patients who received a PARPi, 54% had BRCA2/1 alterations and 16% had no HRR gene alterations. Other alterations include ATM (10%), CDK12 (10%), BARD1 (2%), CHEK2 (7%), and PALB2 (2%). Overall, 37% had PSA50 responses (95% CI 27–47) and 43% had RECIST responses (95% CI 27–61). Patients with BRCA2 alterations had the highest response rates (52% PSA50, 72% RECIST) and biallelic BRCA2 loss was associated with better outcomes. Panel-based mutational signatures (SigMA, SBS3 from DeepSig) were associated with BRCA-mutated status and overall response to PARPi. Six patients without known HRR gene alterations had responses; one was later found to harbor BRCA2 loss. Limitations include the small study size, retrospective design and use of panel-based sequencing.

CONCLUSION

Most PARPi responders had alterations in HRR genes, particularly BRCA2, although some responders lacked known biomarkers associated with PARPi response. Our findings in this real-world dataset support HRR gene testing for PARPi use in mCRPC and highlight the need for novel genome-wide biomarkers for patient selection.

INTRODUCTION

Poly (ADP-ribose) polymerase inhibitors (PARPis) are FDA-approved for the treatment of patients with metastatic castration-resistant prostate cancer (mCRPC) harboring deleterious germline or somatic mutations in homologous recombination repair (HRR) genes. Patients with alterations in genes that are directly implicated in HRR, including BRCA2 and PALB2, are most likely to benefit from PARPis^1–5. The efficacy of PARPis in patients with DNA repair gene alterations indirectly linked to HRR, including CDK12, ATM, CHEK2, remains less clear. While response rates in CDK12-mutated patients are low^6–10, there may be a progression-free survival (PFS) benefit for the combination of enzalutamide and talazoparib¹⁰. For other genes involved upstream in the DNA damage response, including ATM and CHEK2 (6–8% of mCRPC), there is no apparent benefit from PARPis^1,10,11. In addition, some patients with no identified mutations in HRR genes have shown rare responses to PARPi therapy, suggesting the presence of occult molecular changes leading to HRR deficiency^5,12.

Despite these observations, the overall response to PARPis, even among patients with BRCA2 mutations, remains suboptimal, with response rates ranging between 50–60%^6–9. Biallelic BRCA2 gene loss has been shown to correlate with better treatment outcomes^1,13, but beyond the presence of a single deleterious alteration identified on clinical next-generation sequencing, the factors underlying response to PARP inhibition remain unclear. The heterogeneity of the genomic assays incorporated into prospective studies of PARPis has limited the ability to further investigate biomarkers of response. In this study, we analyzed a real-world dataset of mCRPC patients treated with PARPis and investigated the association of genomic features using a single DNA sequencing assay with response to therapy.

MATERIALS AND METHODS

Study design and patients

We analyzed a cohort of mCRPC patients at Memorial Sloan Kettering Cancer Center (MSKCC) who consented to Institutional Review Board (IRB) protocol 12–245 (NCT01775072) and underwent genomic profiling with the MSK-IMPACT assay, which targets 505 cancer-associated genes in its latest iteration^14,15, between July 2014 and April 2023. Germline mutations are reported only for patients who provided specific consent. Clinical data were recorded using a clinical database, for which data definitions and a validation study were previously described¹⁶. Patients were included if they received a PARPi (olaparib, rucaparib, niraparib, or talazoparib) for mCRPC. We excluded cases with poor tumor purity (<1 mutation/MB, <1% genome altered, and <20% cellularity), incomplete baseline data (unknown PARPi start date, no tumor profiling prior to start of PARPi, no baseline PSA), previous platinum chemotherapy or PARPi exposure, and concurrent treatment with another active agent for prostate cancer with the PARPi (Supplementary Figure 1). Radiographic responses via RECIST1.1 were assessed for patients with baseline measurable disease by an experienced radiologist (A.W.). Data collection and analyses were performed with MSKCC IRB approval.

Genomic analysis

Somatic alterations identified were reported as mutated only if classified as ‘oncogenic’ or ‘likely oncogenic’ by OncoKB^17,18. Somatic alterations that were classified as ‘likely passenger’ were not included. Of the patients who consented to germline testing (114/120, 95%), germline variants were reported as mutated only if defined as “pathogenic” or “likely pathogenic” per American College of Medical Genetics (ACMG) guidelines. HRR genes were defined as one of 13 genes directly or indirectly involved in HRR per the PROfound study⁵. Tumor mutation burden (TMB)¹⁹ was categorized as high (≥10 mutations/MB) or low (<10 mutations/MB).

Allele-specific copy number and clonality analysis

Allele-specific copy number and clonality were inferred using the FACETS algorithm (v0.5.14)²⁰.²¹. We applied a series of quality control (QC) criteria (https://github.com/taylor-lab/facets-preview/) and samples that did not pass QC metrics were excluded from this analysis. Clonality of somatic mutations was determined as described previously²¹ (https://github.com/mskcc/facets-suite). Biallelic inactivation was defined as either: homozygous deletion, one oncogenic mutation with concomitant copy-number loss of heterozygosity, or one somatic or germline pathogenic variant paired with a somatic oncogenic mutation in the same gene as previously described²². A total of 37 tumors were excluded from zygosity and clonality analyses due to failing FACETS QC criteria.

Mutational signature analysis

Signature 3 is a mutational signature characterized by a specific pattern of single nucleotide variations (SNVs) and may be a reliable surrogate for HRR deficiency (HRD)²³. Two methods – SigMA²⁴ and DeepSig (https://github.com/mskcc/DeepSig) - were employed to identify the presence of signature 3.

Outcomes analyses

PSA50 response was defined using prostate cancer working group 3 (PCWG3) criteria as a PSA decline of ≥ 50% from baseline, confirmed by a second evaluation at least 4 weeks later²⁵. Baseline PSA ≥ 2.0 ng/ml and at least one PSA during follow-up on treatment were required for patients to be considered evaluable for PSA response. PSA data were collected until December 20, 2023. Radiographic response was defined as a partial response (PR) or complete response (CR) using response evaluation criteria in solid tumors (RECIST, version 1.1)²⁶, in patients with baseline measurable disease. Radiographic data were collected from May 2015 to December 2021. Progression-free survival (PFS) used in the OncoCast analysis was extracted from radiology reports using NLP models as previously described²⁷.

Statistical analysis

Response rates (radiographic or PSA50) were reported in percentage and 95% confidence interval (CI) calculated using the Clopper-Pearson method. Comparisons between continuous values were done by Wilcoxon rank-sum tests, with Dunn’s test for pairwise multiple comparison of ranked data. Statistical analyses were performed using R version 4.3.1.

OncoCast analysis

OncoCast (https://github.com/AxelitoMartin/OncoCast) was used for PFS risk stratification. The lasso method for variable selection was implemented in the R glmnet package. The model integrated genomic alterations, genomic features (TMB, FGA, SBS3 from DeepSig), and clinicopathological variables (Gleason score, baseline PSA, M1 stage) to generate a 0–10 risk score per patient by aggregating prognostic effects across the panel of sequenced genes and clinical data.

Data access

Patient-level genomic data and outcomes are summarized in Supplementary Table 1 and made available at https://www.cbioportal.org/study/summary?id=prad_msk_2025.

RESULTS

Patient characteristics

A total of 120 patients with mCRPC who received a PARP inhibitor were included. Of these, 114 (95%) underwent tumor and germline testing and six had tumor-only profiling. Baseline characteristics are shown in Supplementary Table 2. Among the 120 patients, 65 (54%) had BRCA2/1 mutations, 12 (10%) had CDK12 mutations, 12 (10%) had ATM mutations, 12 (10%) had other HRR gene mutations (2 in BARD1, 8 in CHEK2, 2 in PALB2), and 19 (16%) had no HRR gene mutations (none). Of the 65 BRCA2/1 patients, 21 (32%) had germline mutations and 44 (68%) had somatic mutations. Of the 12 ATM-mutated patients, 6 (50%) had germline mutations. Most patients received olaparib (93), followed by rucaparib (23), niraparib (2), and talazoparib (2). Seventy-one patients (59%) received a prior taxane, and all but one had received a prior androgen receptor pathway inhibitor (ARPI).

Response by gene alteration

Of the 120 patients, 37/120 were evaluable for radiographic response, with 16 (43%; 95% CI, 27–61%) showing a radiographic response. A total of 104 patients were evaluable for PSA response, i.e. had a baseline PSA ≥ 2.0 ng/ml and at least one follow-up PSA on treatment. 38/104 (37%; 95% CI, 27–47%) achieved a PSA50 response (Table 1). Patients with BRCA2/1 alterations had significantly higher PSA50 response rate (53%, p<0.001) and RECIST response rate (62%, p = 0.009) compared with the overall population. CDK12 altered patients had no objective responses (0/12 PSA50 and 0/6 RECIST). ATM altered patients had a PSA50 response rate of 20% (2/10) and RECIST response rate of 33% (1/3). Among patients with other HRR gene alterations, two had PSA50 responses (CHEK2 and BARD1), and one had a RECIST response (BARD1).

Table 1.

PSA50 and RECIST response by HRR gene alteration

	Total	None	BRCA1/2	CDK12	ATM	Other HRR
PSA501
Total with baseline measurable PSA2	111	18	58	12	12	11
PSA50 inevaluable3	7	2	0	0	2	3
PSA50 response rate	38/104 (37%)	3/16 (19%)	31/58 (53%)	0/12	2/10 (20%)	2/8 (25%)6
RECIST4
Total with baseline measurable disease	37	3	21	6	3	4
Response rate (PR or CR)	16/37 (43%)	1/3 (33%)	13/21 (62%)	0/6	1/3 (33%)	1/4 (25%)6
Composite PSA50 or RECIST response rate5	39/115 (33.9%)	3/18 (16.7%)	32/61 (52.5%)	0/12 (0%)	2/12 (16.7%)	2/126 (16.7%)

Abbreviations: PSA, prostate-specific antigen; RECIST, response evaluation criteria in solid tumors.

¹PSA50 response: having a PSA decline of 50 percent or more from baseline.

²Total number of patients with baseline PSA greater than or equal to 2.0 ng/ml.

³No follow up PSA.

⁴RECIST response: complete or partial response as defined by RECIST v1.1.

⁵Rate of PSA50 or RECIST responses among all patients who are evaluable for either a PSA50 or a RECIST response.

⁶The responders in the ‘Other HRR’ category had alterations in BARD1 (Both PSA50 and RECIST responder) and CHEK2.

For patients with BRCA2/1 alterations, the median time on treatment was 6.5 months, with 10 still on treatment at data cutoff (Figure 1). Median time on treatment for patients with alterations in CDK12, ATM, and no HRR gene alterations was 3.5, 3.4, and 2.3 months, respectively. More than 90% of patients were noted to discontinue the PARPi due to clinical or radiographic disease progression, 8% for toxicity. Nearly all patients had a rising PSA before starting PARPi (Supplementary Figure 2). In the CDK12 subgroup, only 1/12 (8%) had a PSA drop of ≥30%. For ATM mutated patients and those without HRR mutations (None), PSA generally increased, though 9 patients (4 ATM, 5 None) had a PSA decline of ≥30%.

Figure 1.

Time on treatment by HRR gene alteration

Time on treatment in weeks for patients with no HRR alteration, BRCA2/1, ATM and CDK12 alterations. The dotted line indicates the median time on treatment for each group. The star at the end of the bar denotes patients with an objective response (PSA50 or RECIST).

Responses for patients with BRCA2/1 alterations

Among the 65 patients with BRCA2/1 alterations, 58 were evaluable for PSA50 response, and 21 for radiographic response. The radiographic response rate for BRCA2 patients was 13/18 (72%; 95% CI, 47–90%), while none of the 3 BRCA1 patients had a RECIST response (Figure 2.). The PSA50 response rate was 27/52 (52%; 95% CI, 38–66%) for BRCA2 and 2/6 (33%; 95% CI, 4–78%) for BRCA1. Response rates were similar for somatic and germline BRCA2/1 alterations and no clear difference in response rate was observed by type of alteration (e.g. mutation versus deletion). Clonality analysis of BRCA2/1 mutations did not identify sub-clonal BRCA mutations, and response rates for cases with confirmed clonal mutations in BRCA2/1 were similar to the overall cohort.

Figure 2.

RECIST and PSA50 responses for patients with BRCA2 and BRCA1 alterations

The vertical dotted lines indicated the radiographic objective response rate (ORR) and PSA50 response rate in the overall population of evaluable patients with BRCA2/BRCA1 alterations. None of the patients with evaluable clonality were found to have a sub-clonal mutation in BRCA2/1.

Patients with confirmed biallelic alterations in BRCA2/1 had a better response to PARP inhibition (ORR 77%; 95% CI, 46–95%; PSA50 61%; 95% CI, 41–78%) than patients with monoallelic alterations (ORR 0%; 95% CI, 0–71%; PSA50 31%; 95% CI, 9–61%)) (Figure 2 and Supplementary Figure 3). Patients with TMB-H tumor had limited responses to PARP inhibition with only 1/3 (33%) achieving a PSA50 response.

No significant difference in response was observed for patients with BRCA2/1 alterations who had a co-occurring TP53, PTEN, RB1 or ERG alteration. However, patients with PTEN alterations had a numerically higher response rate (ORR 83%, 95% CI 36–100%; PSA50 64%, 95% CI 31–89%), while patients with RB1 alterations tended to have fewer responses (ORR 0/2, PSA50 29%, 95 %CI 4–71%).

Arm level gains or losses from panel sequencing were analyzed in 61 patients with BRCA2/1 alterations by response (PSA50 or RECIST). The most frequent CNAs were losses in 8p, 13q, and 16q, and gains in 8q, consistent with previous studies in prostate cancer. No significant correlation was found between response and arm level copy number gains or losses, including 10q which harbors PTEN, 13q which harbors RB1, and 17p which harbors TP53 (Supplementary Figure 4).

Mutational signatures and response to therapy

Both SigMA and DeepSig SBS3 scores from our panel sequencing data were used to identify signature 3. Both SigMA scores and DeepSig SBS3 scores were significantly higher in BRCA2/1-altered tumors than in the other subgroups (Figure 3a, 3c), and also higher among RECIST and/or PSA50 responders compared to non-responders in the overall cohort (Figure 3b, 3d), but not significantly different between BRCA2/1 responders and non-responders.

Figure 3.

Mutational signature analysis by gene alteration, and association with response

a) SigMA score by HRR gene alteration. None = no HRR gene alteration. Others = other HRR gene alteration b) SigMA scores for responders (PSA50 or RECIST) and non-responders to PARPi. c) SBS3 score from DeepSig by HRR gene alteration. d) SBS3 scores from DeepSig for responders (PSA50 or RECIST) and non-responders to PARPi.

PARPi response prediction using LASSO model

Using OncoCast, we calculated risk scores based on clinical and genomic data to identify patients at higher risk of progression (n=115). Patients were stratified into two risk groups. Low PSA correlated with longer PFS, while TP53, AR and CDK12 somatic alterations linked to poorer outcomes (Figure 4a, 4b). Other factors associated with worse PFS included Gleason score >8 and PSA >100. The high-risk group had a median progression time of 2.2 months (95% CI: 1.8–3.0), while the low-risk group had 5.5 months (95% CI: 2.8 – 9.7) (Figure 4c).

Figure 4.

PARP response prediction using LASSO model

a) Heatmap for visualizing each binary feature as a function of risk score. The color blue represents presence and grey represents absence. Features with more blue lines to the right of the plot drive the risk score upward (unfavorable) and those to left of the plots drive downward (favorable). (b) Volcano plot showing TP53, AR, CDK12, Gleason score >8 and PSA above 100 are associated with worse radiographic progression-free survival (PFS) (pink color, right side of vertical line). Other features associated with worse PFS include M1 stage, TMPRSS2-ERG fusion (green dots). PSA (0–4) is associated with better PFS (pink color, left side of vertical line). (c) Kaplan-Meier plot of PFS for the two risk subgroups.

Non-BRCA responders to PARP inhibition

Of the 115 patients who were evaluable for either PSA50 or RECIST response, 39 (33.9%) were responders (Table 1). Among them, 32 (82%) had BRCA2/1 alterations, 3 had no HRR gene alteration, 2 had ATM alterations and 2 had other HRR gene alterations (BARD1 and CHEK2). In total, 33/39 responders (85%) had an alteration in BRCA2/1 or another gene directly implicated in HRR.

Six patients had an objective response, but did not have an alteration in a gene with a direct role in HRR (2 with ATM, 1 with CHEK2, 3 with no HRR gene alteration) (Figure 5). Among them, one germline ATM-altered patient had exceptional response to olaparib, with a 98% PSA decline and remained on treatment for over 20 months (Figure 5a). This patient exhibited a Signature 3 profile with a SigMA score significantly higher than most patients with ATM alterations, surpassing the median observed in BRCA2/1-mutated patients (Figure 5b). Subsequent tumor profiling revealed a BRCA2 deletion not found in the pre-treatment sample.

Figure 5.

Analysis of PARP inhibitor responders without a direct HRR gene mutation

a) In total, 6 of the 39 responders to PARP inhibition did not have an alteration in BRCA2/1 or another gene directly implicated in HRR. Of those, 2 had mutations in ATM (dark gray bars), 1 in CHEK2 (light gray bar), and 3 had no DDR gene mutation (black bars). Shown here is their PSA change from baseline, PSA50 response (blue), both PSA50 and RECIST response (red), germline mutational status, co-occurring gene alteration, and the presence of signature 3. b) SigMA scores for the 6 unexplained responders relative to SigMA scores for all patients with BRCA2/1, ATM, CHEK2, and no HRR alterations (None). The two patients with ATM mutations are represented with red dots, the one patient with CHEK2 mutation is represented with blue dot, and the patients without HRR gene alterations are shown in green dots. One responder without an HRR gene alteration had an unknown SigMA score and is not shown here.

One patient without an HRR alteration showed a remarkable PSA response (99% decrease) and remained on treatment for over 15 months (Figure 5a). Although the patient declined germline testing despite a family history of pancreatic and breast cancer suggesting a possible inherited HRR gene mutation, the tumor showed no evidence of Signature 3. The other 4 non-BRCA responders also lacked Signature 3, and their responses remain unexplained.

DISCUSSION

Several prospective phase 2 and 3 studies have reported outcomes for PARPi-treated mCRPC patients with or without HRR gene alterations. Our real-world cohort of 120 mCRPC patients treated with PARPis included 58% with alterations in HRR-specific genes (BRCA2/1, BARD1), 26% with alterations in genes indirectly implicated in HRR (ATM, CHEK2, CDK12), and 16% with no known HRR mutations. We evaluated response as ORR based on RECISTv1.1 and PSA decline ≥50% (PSA50). Among the 115 evaluable patients, the overall response rate (ORR or PSA50) was 34% (95% CI, 25–43%). Of responders, 85% had an HRR-associated mutation (77% BRCA2, 5% BRCA1, 3% BARD1), and 8% had no HRR mutation. These findings confirm that most PARPi responders can be identified by the presence of a deleterious HRR gene alteration, such as BRCA2/1, BARD1, or PALB2.

The response rate for BRCA2/1 in our real-world cohort was comparable to that in prospective phase 2 trials (TOPARP-B, TRITON2, GALAHAD, TALAPRO-1)^1,7–9. Although radiographic progression-free survival (rPFS) could not be assessed due to the heterogenous imaging follow-up, the median time on PARPi therapy for BRCA2/1 patients was unexpectedly shorter at 6.5 months, compared with a median PFS of approximately 9 months in prospective trials. This short duration on therapy was not explained by more aggressive or advanced disease in our cohort and may reflect clinical practice patterns, where physicians often switch therapy earlier in standard practice compared with trial settings – a phenomenon also observed with ARPIs²⁸.

Among patients with alterations in BRCA2 and BRCA1, biallelic loss was associated with better response, as previously reported^1,13. We assessed whether BRCA2/1 mutation clonality predicted response but found no sub-clonal BRCA mutations, supporting that BRCA2 deleterious mutations are predominantly truncal oncogenic events early in the prostate cancer pathogenesis^29–31. One limitation is that 37 tumors were excluded from zygosity and clonality analysis due to failing quality metrics for this type of analysis, although the possibility of misclassification still exists. We did not include in this analysis patients with BRCA mutations identified on plasma DNA testing, where sub-clonal alterations may be more prevalent. We found no other co-occurring genomic alterations significantly associated with response to PARPis, although co-occurring RB1 alterations showed a trend toward worse response. HRD-associated panel-based mutational signatures were strongly associated with BRCA alterations and with response to PARP inhibition in our study, as previously reported for prostate cancer and other malignancies^32–34. However, panel-based mutational signatures did not differentiate between BRCA2/1 responders and non-responders.

Consistent with previous studies, a minority of 6 (15%) PARPi responders in our cohort had no known HRR gene mutation, or had a mutation in ATM or CHEK2, genes indirectly implicated in HRR with low PARPi response. One patient with a germline ATM mutation had a PSA decline of > 90% and remained on treatment > 20 months, with a high HRD signature score and a BRCA2 deletion identified in a post treatment tumor sample. This patient’s response likely resulted from baseline BRCA2 loss not detected in the pre-treatment sample due to low tumor purity. The remaining 5 responders had no discernable explanation for their response. Panel-based mutational signature analysis did not predict PARPi responders beyond the presence of a deleterious HRR mutation. No benefit was observed in the 12 patients with CDK12 mutations, who had no objective responses, limited treatment duration, and unchanged PSA trends.

Our study provides the largest real-world dataset of mCRPC patients treated with PARPis and profiled using a single genomic assay. We conclude that most responders can be identified by the presence of a deleterious HRR gene mutation, primarily BRCA2, but that some patients who benefit lack the standard biomarkers used for PARP inhibitor selection. Limitations include the study’s size, the use of a panel-based NGS assay, and the study’s retrospective design. In particular, panel-based mutational signatures may not capture genome-wide genomic changes associated with HRD and with benefit from PARP inhibition. Larger studies with advanced molecular profiling, including expression profiling and whole genome sequencing, may be necessary to improve biomarker selection for PARP inhibitor therapy in mCRPC.

CONTEXT SUMMARY.

Key objective:

What are the genomic features associated with response to PARP inhibitors in metastatic castration-resistant prostate cancer?

Knowledge generated:

The majority of PARP inhibitor responders carry a somatic or germline mutation in a gene involved in homologous recombination repair (HRR) and a mutational signature associated with homologous repair deficiency (HRD). Some patients who benefit have no genomic evidence of HRD on panel sequencing.

Relevance:

Most patients with mCRPC who respond to PARP inhibitors can be identified by the presence of an HRR gene mutation on panel sequencing, but whole genome/transcriptome analyses will likely be needed to identify all patients who may benefit.