HRDetect in Ovarian Carcinoma: Stratification and Therapeutic Implications

Distribution and Implications of HRDetect in UW OC cohort

All patients in the cohort had undergone maximal effort surgical cytoreduction, and all but one received SOC chemotherapy with a platinum-doublet. Ten patients received neoadjuvant chemotherapy. WGS was performed on samples taken prior to treatment in 176 cases and from post-treatment recurrences in 9 samples (Supplementary data Fig1). Similar to TNBCs previously, more than half the cancers were HRDetect-high (95/185, 51.4% - exceeding a predefined score of 0.9, predictive of HRD), which correlated with a significantly improved overall survival (OS) (median OS 6.2 vs 4.1 years, HR=0.6, 95% CI=0.41-0.87, log-rank p=0.007) (Fig1a,h). A remarkable finding was that 23.2% (43/185) were HRDetect-intermediate (score 0.1-0.9), in contrast to only ∼13.5% of TNBCs². Another 25.4% (47/185) were classified as HRDetect-low (score <0.1) (Fig 1a). Although HRDetect prognostically distinguished HRDetect-high from intermediate or low, (median OS 6.2 vs 4.8 vs 3.4 years; log-rank p = 0.02) (Fig 1i), there was no survival difference between individuals with HRDetect-intermediate and -low OC, out-of-keeping with TNBC where HRDetect-intermediate cases had the poorest outcomes².

• Download figure
• Open in new tab

Figure 1: Clinical and genomic characteristics of HRDetect classification in the UW ovarian cancer cohort.

A) Bar plot of HRDetect probability in 185 OCs b) genetic alterations in homologous recombination associated genes and other drivers c) d) & e) Proportions of mutational, indel patterns, and rearrangement signatures are shown as proportional bar plots, f) Proportions of rearrangement signatures as a heat map g) Rearrangement clusters of each case h) Kaplan–Meier analysis of association of OS with binary HRDetect classification i) Kaplan–Meier analysis of association of OS with HRDetect-high versus HRDetect-intermediate versus HRDetect-low j) Kaplan–Meier analysis of association of OS with rearrangement clusters 4&6 versus 5&8 versus 9-11 k) Correlation between genomic LOH and HRDetect scores.

Genomic Features Associated with HRDetect Categories

We sought causes of high HRDetect scores. Of 95 HRDetect-high cases, 46 (48.4%) had deleterious germline and/or somatic mutations in BRCA1 or BRCA2 with concomitant loss of the wildtype allele (Fig 1a and Supplementary data Fig 5). Promoter hypermethylation was observed in BRCA1 in 21 (22.1%) and in RAD51C in one (1%). Pathogenic germline variants were observed in other key HRD genes in 6 cases (6.3%); 2 in RAD51C, one in RAD51D, one in PALB2, and two in BRIP1, all with loss of the other allele. The causes for HRDetect-high scores in the residual 21 (22.1%) remain unexplained, although 4 OC had BRCA1 promoter hypermethylation with LOH, albeit below the predefined quantitative cut-off of 60% methylation⁴. Notably, these 21 HRDetect-high cases would have been missed using conventional targeted-sequencing panels and gene-specific epigenetic assays. While the application of customary copy-number-based gLOH assay obtained from WGS would have captured 93 of 95 cases, it will also have called an additional 62 of 90 remaining cases as HRD, suggesting lower specificity for gLOH(Fig1k)⁵.

We also asked whether there were prognostically differentiating driver mutations in the UW cohort (Fig1b, Supplementary data Fig 3b). 19/21 (90.4%) of the cases had which RB1 driver mutations were enriched amongst HRDetect-high platinum-sensitive cases (p=0.0005). Interestingly, the singular RB1 mutant case that had platinum-refractory disease (progression during primary adjuvant chemotherapy) had a HRDetect-low score. 29 OC had CCNE1 amplifications, of which two were platinum-refractory. CCNE1 amplifications were underrepresented in HRDetect-high cases and equally frequent in HRDetect-intermediate and -low categories. Each had another driver event - a MYC amplification and a CDK12 mutation respectively. All other drivers were not adequately powered for a clear analysis but were without clear prognostic impact in this cohort. 14 OC had MYC amplifications and distributed across all three HRDetect categories with 4, 5, and 5 cases in the high, intermediate, and low respectively. Two of these cases were platinum-refractory and had another driver event – a PIK3CA mutation and the CCNE1 amp mentioned above. Four OC had CDK12 drivers. One was HRDetect-high and platinum-sensitive, two were HRDetect-intermediate, one platinum-sensitive, and one platinum-refractory, and one was HRDetect low with unknown platinum status. There were also four OC with PIK3CA mutations (two HRDetect-high and two low).

Rearrangement Signatures as Classifiers and Prognosticators Beyond HRDetect

In clinical settings, OC is often dichotomously defined as HRD or HR proficient (HRP). Critically, we find that the ‘HRP’ category is highly heterogeneous, enriched with distinguishing rearrangement signatures (RS) including RS1, RS14 (long tandem duplicator phenotypes), R2, R6a, and R6b (characterized by complex, focal amplifications), and R18 (defined by inter-chromosomal translocations) (Fig 1 a,e,f, Supplementary data Fig 3). To independently evaluate the importance of these RS, we performed unsupervised hierarchical clustering using only RS and evaluated the OS of resulting clusters. Eleven clusters were identified (Fig 1 g, Supplementary data Fig 4). Focusing on clusters that had greater than 10 samples, clusters associated with HRD (clusters 9, 10, and 11) were characterized by RS3 and RS5 and predominantly had very high HRDetect scores (74/87 >= 0.98)). They mirrored HRDetect prognostication and most had favorable outcomes (Figure 1j). Cluster 4, characterized by a high proportion of RS2, and cluster 6, with high proportion of RS1, were associated with the poorest outcomes, while the intermediate outcome clusters contained those with relatively quiet genomes (cluster 8) and cluster 5, containing a mixture of samples with both RS2 and RS3. These results suggest that in the HRP category, rearrangement signatures may serve as discriminating biological signals, and can make a further contribution towards prognostication, beyond HRDetect scoring alone.

Characterizing HRDetect-Intermediate and -Low OC

To better understand this, we evaluated the 90 HRP cases in more detail. First there were commonalities between the 43 HRDetect-intermediate (probability score between 0.1 and 0.9) and 47 HRDetect-low cases (score <0.1): 30 of 43 (69.8%) and 30 of 47 (63.8%) cases of HRDetect-intermediate and -low respectively, carried long TD phenotypes of RS1 or RS14 and/or a substantial contribution from RS18 (translocations), compared to 14/95 (14.7%) of HRD (HRDetect-high (score >0.9)). The strongly negative prognostic TD phenotypes are prevalent in both intermediate and low groups, hence, survival analyses contrasting HRDetect-intermediate and -low are non-discriminating. Of the remaining 13/43 HRDetect-intermediate cases, 10 had R2-related translocations or R7-long deletions (>10kb). Two were chromosomally quiet genomes. Amongst HRDetect-low cases, of the remaining 17/47 cases, 8 had RS2, RS6a or 6b. Nine were chromosomally quiescent, and one case had a striking mismatch repair deficiency, with the typical substitution (SBS26 characterized by T>C transitions) and indel features (small insertions at poly nucleotide repeat tracts) caused by PMS2 abrogation, a known cause of mismatch repair deficiency as confirmed through both cancer analyses ⁶and in human experimental systems ⁷.

Predictive Value of HRDetect in the ARIEL2 Cohort

To next assess whether HRDetect could be predictive of sensitivity to PARPi, we performed WGS in paired tumor-normal DNA from ARIEL2, a one-arm phase II treatment trial of rucaparib in recurrent, platinum-sensitive, high-grade OC with measurable disease. 77 cancers were evaluated by WGS. 37(48%) were HRDetect-high, 12(15%) HRDetect-intermediate, and 28(37%) were HRDetect-low (Fig 2a). The proportions in the ARIEL2 cohort are different from the UW cohort, reflecting differential selection for relapsed, platinum-sensitive OC in the context of clinical trial.

• Download figure
• Open in new tab

Figure 2: Clinical and genomic characteristics of HRDetect classification in patients treated with rucaparib in the ARIEL 2 clinical trial.

a) Bar plot of HRDetect probability in 77 OCs on clinical trial b) genetic alterations in homologous recombination associated genes and other drivers, clinical responses to PARPi therapy c) & d) Proportions of mutational and rearrangement signatures are shown as proportional bar plots, e) Proportions of rearrangement signatures as a heat map f) Kaplan–Meier analysis of association of PFS with binary HRDetect classification g) Kaplan–Meier analysis of association of PFS with genomic LOH h) Kaplan– Meier analysis of association of PFS with HRDetect-high versus HRDetect-intermediate versus HRDetect-low i) Correlation of genomic LOH with HRDetect scores.

Patients with HRDetect-high OC treated with PARPi had a significantly improved progression-free survival (PFS) compared to the rest (median PFS 11.1 vs 7.1 months, HR=0.44, 95% CI=0.26-0.74, log-rank p=0.03) (Fig 2f) and treatment overall response-rates (ORR=54% vs 22.5%), demonstrating that HRDetect is an excellent identifier of HRD and predictor of therapeutic response to PARPi in OC. Causes of HRD were identified in 56.8% of HRDetect-high cases including 15 (40.5%) with germline or somatic mutations in BRCA1/BRCA2, one germline mutation in each of RAD51C, BRIP1, and RAD51D, and three with BRCA1 promoter hypermethylation (Fig 2b). The causes for HRDetect-high scores in the remaining 16 (43.4%) samples were unclear. Enrollment criteria for ARIEL2 enriched for BRCA wildtype cases ⁸, explaining the higher fraction of unexplained high HRDetect in this cohort.

Comparison with Genomic LOH HRD and Implications for Clinical Use

Finally, we compared with genomic LOH HRD as defined by FoundationOne™ NGS assay used in ARIEL2. There was a modest correlation R² = 0.34 (Fig 2i) with 29 (78.4%) HRDetect-high cases also having high LOH. 7 (17.5%) of HRDetect-low/intermediate had high LOH. 6 cases were called LOH high but did not have a high HRDetect score nor carried mutations to explain reported LOH. Median PFS for LOH high vs low cases was 11.3 vs 5.5 months (HR= 0.49, 95% CI = 0.29-0.87, log-rank p=0.0017) (Fig 2g) as a predictor on PARPi therapy, contrasting the curves between HRDetect and LOH did not reveal a significant difference between them.

In summary, our study demonstrates that HRDetect prognosticates effectively and accurately predicts HRD and PARPi response in OC. Beyond HRDetect, WGS reveals a continuum of biology and survival in non-HRD OC. Key indicators of improved response to SOC include a very high HRDetect score (>0.98) and RB1 mutations. Of other HRDetect groups, CCNE1 amplifications could help improve prognostic status. However, high RS1, RS14, and RS18 are likely indicators of poor outcomes and alternative therapeutic strategies must be considered for these patients. Crucially we now have a way of stratifying these cancers using WGS strategies.

Ethics approval and consent to participate

Ethics approval and consent to participate. The UW Ovarian Cancer Biorepository was approved by the Institutional Review Board (IRB). All patients provided written informed consent prior to enrolment. Likewise, ARIEL2 participants provided informed consent, including for translational studies.

University of Washington (UW) Patient cohort

Patients diagnosed with high-grade OC, undergoing surgery, at UW were prospectively enrolled in a biorepository. Of these, 212 representative patients with 222 samples were selected for this study (Supplementary fig 1). 31 samples were eliminated from analysis for quality control reasons leaving 191 neoplastic samples from 185 patients. There were 170 (92%) patients with samples collected at primary debulking surgery only, 9 (5%) patients with samples collected at recurrence only, and 6 patients (3%, 12 samples) with samples collected at both presentation and recurrence. All patients were followed until death and, if alive, censored at the last follow-up.

ARIEL2 cohort

ARIEL2 part 1 was an international, multicenter, phase 2, open-label study of oral rucaparib in patients with recurrent, platinum-sensitive, high-grade ovarian carcinoma, which were classified into one of three predefined homologous recombination deficiency subgroups based on tumor mutational analysis: BRCA mutant (deleterious germline or somatic), BRCA wild type and LOH-high (LOH-high group), or BRCA wild type and LOH-low (LOH-low group). A cutoff of 14% or more genomic LOH was prespecified for the LOH-high group ⁸.

Tissue sampling, DNA, and RNA extraction

Fresh tumor samples preserved in RNAlater (Qiagen) were obtained with routine clinical sampling. RNA and DNA were extracted using the Qiagen Allprep extraction kit (Qiagen) as previously described. ARIEL2 included DNA extraction formalin fixed paraffin tissues (FFPE) as previously described ¹⁹.

Whole genome sequencing analysis

DNA extracted from both the tumor and corresponding normal tissue were subjected to whole-genome sequencing to an average sequencing depth of 29X in the tumor and 27X in the matched normal on Illumina sequencers. The resulting BAM files were aligned to the reference human genome (GRCh37) using Burrows-Wheeler aligner, bwa mem 0.7.17-r1188 (dockstore-cgpmap 3.0.4 https://quay.io/repository/wtsicgp/dockstore-cgpmap. Mutation calling was performed as described previously^{10,11 .19,}The algorithms used are contained in the following dockstore,(dockstore-cgpwgs v2.1.0 https://quay.io/repository/wtsicgp/dockstore-cgpwgs).

Mutational signatures and HRDetect

Mutational signatures were assigned using FitMS 2.4.1 ²¹. For substitution signatures, a two-step approach was used, first fitting a set of reference signatures commonly found in ovarian cancer, followed by rare signatures previously found in ovarian cancer. This two-step approach improves the solution in cases where substantial residual mutations remain after fitting the common signatures, whilst minimizing inappropriate overfitting of mutational signatures. HRDetect is a weighted model that combines the contributions of six distinguishing mutational signatures to produce a signal score that is predictive of a defect in the homologous recombination DNA repair pathway¹. Following signature assignment, the contributions of substitutions signatures SBS3 and SBS8, rearrangement signatures RefSig R3 and RefSig R5, HRD LOH index and proportion of small deletions at regions of microhomology were used to calculate the HRDetect probability score. Code for the versions of both FitMS and HRDetect used can be found here (https://github.com/Nik-Zainal-Group/signature.tools.lib.)

Hierarchical clustering

Column-based hierarchical clustering of rearrangement reference signatures was performed with a default setting of pheatmap package (Version 1.0.12) and cutree_cols setting was set as 10. Visualization plots were created with ggplot2(2.2.1) and custom R scripts.

Methylation analyses

Neoplastic DNA was bisulfite converted using the EZ DNA Methylation-Lightning kit (Zymo Research). Droplet digital PCR was performed using the QX200 AutoDG System (Bio-rad laboratories). Primers were designed targeting a 72-bp amplicon in the BRCA1 promoter region associated with transcriptional repression, with specific QSY hybridization probes targeting the fully methylated (FAM-labeled) and the fully unmethylated sequences (VIC-labeled). Quantitative methylation (fraction of methylated droplets divided by all positive droplets) was corrected for neoplastic cellularity from whole genome sequencing estimation. Samples with ≥ 60% methylation were considered positive for methylation.

Methylation analysis on samples from ARIEL2 was done as previously described ²²

Statistical analyses

Survival analyses and figures were created in GraphPad Pris,(version 10.0.2). Survival curves were compared using Kaplan–Meier estimates and the log-rank test. Hazard ratios were calculated through univariable or multivariable Cox regression using the coxph R function. Statistical comparisons between groups were performed using Wilcoxon’s or Kruskal–Wallis tests for numerical values or the chi-squared test for ordinal values. All P values reported from statistical tests are two-sided if not otherwise specified.