Validation of the Clinical Use of GIScar, an Academic-developed Genomic Instability Score Predicting Sensitivity to Maintenance Olaparib for Ovarian Cancer Open Access

PARP inhibitors have transformed the management of high-grade serous ovarian carcinoma, consistently demonstrating remarkable efficacy in randomized clinical trials in both newly diagnosed and platinum-sensitive recurrent disease, with HRs for progression-free survival (PFS) ranging from 0.30 to 0.59 (1–8). With the integration of PARP inhibitors into routine clinical practice, the determination of homologous recombination deficiency (HRD) status has become increasingly important in therapeutic decision-making, especially in first-line decisions for which clinical indicators of PARP inhibitor sensitivity such as platinum-free interval are not available. SOLO-1, the first trial to show a PFS benefit from maintenance PARP inhibitor therapy after response to first-line chemotherapy, restricted enrollment to patients whose tumors harbored BRCA1/2 mutations (1). However, subsequent trials in broader populations indicated that the population benefiting from maintenance PARP inhibitors after first-line chemotherapy for ovarian cancer extends beyond those with BRCA-mutated disease. In the PAOLA-1 trial, the addition of maintenance olaparib provided a significant PFS benefit, which was substantial in patients with HRD tumors (hereafter described as HRD-positive), including those without a BRCA mutation (2). In the PRIMA trial, benefit from niraparib expanded further to patients without HRD (also sometimes described as homologous recombination proficient and hereafter termed HRD-negative), although the PFS benefit was more modest in HRD-negative populations (3, 9). On the basis of these results, European guidelines recommend that HRD testing is done before the end of first-line chemotherapy, following or preferably together with BRCA testing (10). More recently, long-term overall survival (OS) data from PAOLA-1 have raised questions about the value of maintenance PARP inhibition after first-line chemotherapy in HRD-negative ovarian cancer (11); therefore, robust and reliable HRD testing remains essential in everyday management of ovarian cancer. Currently, there is no gold standard to determine HRD status except the clinical response to dedicated treatment.

In both PAOLA-1 and PRIMA, HRD status was determined using the commercially available Myriad Genetics MyChoice CDx (MGMC; Myriad Genetics), which detects tumors harboring either pathogenic variants in BRCA1 or BRCA2 genes or a genomic instability score (GIS) above the threshold of 42 (12–16). GIS calculation is based on LOH, telomeric allelic imbalance, and large-scale transitions. The estimation of these three values is based on the sequencing of >55,000 SNP loci spread across the human genome. The alternative test commercially available as a companion diagnostic HRD test is the FoundationOne CDx next-generation sequencing (NGS) assay (Foundation Medicine Inc). Briefly, this test is based on the sequencing of >3,500 SNP evenly distributed across the genome and quantifying the extent of genomic LOH (17, 18). This comprehensive genomic profiling test, which was used in the more recently reported ATHENA-mono trial (4), detects alterations in homologous recombination repair (HRR) genes, genomic signatures including microsatellite instability, tumor mutation burden, and LOH.

Despite the regulatory approval of these tests as companion diagnostics for PARP inhibitors in the United States, in many countries, access to commercial HRD tests is an issue and/or expensive, with lengthy turnaround times and high failure rates. Therefore, there is considerable demand for access to cheaper, more readily available tests to determine HRD status and thus guide treatment decisions (19–22). In this article, we present the first French academic solution to HRD testing, called Genomic Instability Scar (GIScar). GIScar was designed to enable identification of patients who can derive benefit from maintenance therapy with the PARP inhibitor olaparib. Our approach was based on targeted DNA sequencing using a 127-gene panel in formalin-fixed paraffin embedded (FFPE) tumors. The GIScar test used these sequencing data to determine a GIS (Fig. 1A). It was developed using a training set of 250 ovarian cancers and a historical commercial MyChoice test provided by Myriad Genetics. Subsequently, the GIScar test was validated on 469 DNA tumor samples from patients treated in the PAOLA-1 clinical trial evaluating the addition of olaparib to maintenance bevacizumab following response to first-line chemotherapy and bevacizumab (2, 23). Here we describe the development of GIScar and its performance in predicting sensitivity to olaparib in terms of PFS and overall survival (OS).

In the first step, between 2020 and 2022, our laboratory (Centre François Baclesse, Caen, France) prospectively collected 131 archival FFPE tumor tissues from patients with ovarian cancer. All patients signed a consent form for the retention, use, and transfer of samples for scientific research purposes. This study was noninterventional, and no additional information or samples were requested from patients, so additional Institutional Review Board approval was not required. An additional 119 FFPE tumor tissue samples from diagnosis of ovarian cancer were collected from other laboratories. Of these, 24 FFPE tissue samples collected by the Institut de Cancérologie de l'Ouest (Paul Papin, Angers, France) were sequenced using our NGS platform. The robustness of our approach was assessed by integrating the remaining 95 FFPE tumor tissues analyzed by NGS from three external platforms into our collection. These comprised 32 FFPE tumor tissue samples from the Département de Génétique, Hôpital Pitié-Salpêtrière (Paris, France), 15 from the Service de Génétique des Tumeurs, Gustave Roussy (Villejuif, France), and 48 from the Unité d'Oncologie Moléculaire Humaine, Centre Oscar Lambret (Lille, France). The overall collection of 250 samples with HRD status determined by MGMC constituted the prospective collection used to train and validate the GIScar model (Fig. 1B; Supplementary Table S1).

In the second step, to assess the predictive value of GIScar in patients treated with olaparib, we used a collection of 469 DNA tumor samples from patients treated in the PAOLA-1 clinical trial (NCT02477644; ref. 2), prepared by investigators from ARCAGY-GINECO (Fig. 1C; Supplementary Table S2). For each patient, HRD status was established by MGMC. We used the mutation status of genes common to the MGMC gene panel and our gene panel to check patient identity. This clinical collection was used to assess whether GIScar results correlated with PFS (assessed as the primary endpoint of the PAOLA-1 trial in patients receiving maintenance olaparib).

Overall, we collected 719 DNA tumor samples (250 + 469) for this study. The DNA tumor samples were from high-grade serous ovarian cancer for 91.0% (654/719; Supplementary Table S3). The histologic type for the other ovarian cancer were endometrioid ovarian cancer (3.2%, 23/719), clear cell ovarian cancer (1.0%, 7/719) and low-grade serous ovarian cancer (0.6%, 4/719). For 31 ovarian cancers (4.3%, 31/719), the pathologist was unable to determine the histologic type.

We started the sequencing from 100 to 200 ng of DNA extract by a Maxwell automate or by a Qiagen kit (QIAamp DNA FFPE Tissue, catalog No. 56404). DNA was fragmented using the Covaris E220 (Covaris). The fragmented DNA was processed by the SureSelect XT HS2 (catalog No. G9983A; Agilent). A custom panel of 127 genes was considered for the enrichment of regions of interest (Supplementary Table S4). Our panel includes the 15 HRR genes (BRCA1, BRCA2, ATM, BARD1, BRIP1, CDK12, CHEK1, CHEK2, FANCL, PALB2, PPP2R2A, RAD51B, RAD51C, RAD51D, RAD54L) sequenced by MGMC. In addition to these 15 HRR genes, 14 genes were included as their alteration leads to therapeutic decision-making. Seventy genes were added for clinical trials (GREAT: NCT04027868, PRO-CARBO: NCT03652493, UTOLA: NCT03745950), and 65 genes were considered as suspected involved in the HRR pathway. The homemade design capture probes were provided by Agilent. After normalization, the libraries were sequenced on a NextSeq 500 device (Illumina) with 16 samples per 2×75 mid-output flowcell (catalog No. 20024904) or 48 samples per 2×75 high-output flowcell (catalog No. 20024907).

The fastq files were generated from the raw data of sequencing by Bcl2Fastq v2.20 (RRID:SCR_015058) or by BCL convert v3.8.2, tools provided by Illumina. Alignment was performed by BWA v0.7.17 (RRID:SCR_010910), using the genome assembly version GRCh37/hg19. PCR duplicates were removed by Picard (RRID:SCR_006525) v2.21.7. HaplotypeCaller embedded in GATK v4.1.6.0 (RRID:SCR_001876) performed the variant calling. The calling of genomic events, such as copy-number variations, was done by CNVkit pipeline v0.9.7 (RRID:SCR_021917). A baseline from analysis of five no-tumor FFPE tissues was constituted as reference for CNVkit running. As the CNVkit needs an estimation of tumor purity, we developed a homemade R program to adjust the purity given by the pathologists. This program was based on the Allele-Specific Copy-number Analysis of Tumor (ASCAT) algorithm (24). Briefly, this approach estimates the tumor ploidy and purity from the log₂ coverage depth and the B-allele frequency of genomic segments. These segments were estimated by the segmentation step of CNVkit. The final outcome was the final calling of CNVkit with segments log2 coverage depth and B-allele frequency adjusted according to the newly estimated purity. Samples were kept for analysis only if ≥50% of targeted regions were sequenced with >100× of coverage depth and if the tumor purity was above 20%. We hypothesized that sequencing data from samples that did not meet these criteria would be insufficient for performing GIScar scoring.

Three variant callers were used from the aligned data by genome-wide association: HaplotypeCaller, LoFreq (25), and OutLyzer (26). The variants were annotated by ANNOVAR tool (27). The variants were interpreted according to the French Genetics and Cancer Group (https://recherche.unicancer.fr/fr/les-groupes-d-experts/groupe-genetique-et-cancer/). Only pathogenic or likely pathogenic variants were reported.

The output from this model is a mathematical Esperance |${\rm{{\rm E}}}( X )$|⁠. An HRD-negative tumor should have a null Esperance |${\rm{{\rm E}}}( {{\rm{HR}}{{\rm{D}}}_{{\rm{negative}}}{\rm{\ }}} )\ = \ 0$| and an HRD-positive tumor should have an Esperance close to 1 ( |${\rm{{\rm E}}}( {{\rm{HR}}{{\rm{D}}}_{{\rm{positive}}}{\rm{\ }}} )\ = \ 1$|⁠). The optimal cutoff was defined on our prospective collection (n = 250 sample) according to MGMC classification by receiver-operating characteristic (ROC) analysis to obtain the optimal value for sensitivity and specificity.

Model training and statistical analyses were done using the software R v3.5.1 (RRID:SCR_001905). We used the library ROCR for ROC analysis and survival library for the time-to-event analysis. The agreement between MGMC and GIScar tests was also evaluated by the Cohen's kappa statistics, the agreement proportion corrected by hazard versus maximum agreement that can be observed (28). The weight of each initial variable (LGE, SIS, and AI) in the GIScar model was assessed by a stepwise approach using the Wald test to discriminate variables between them. For the time-to-event analysis, we focused on PFS and OS from the PAOLA-1 clinical trial. Log-rank tests were used to compare PFS between groups. HRs were calculated from a semiparametric Cox model test.

The tumors used for GIScar training are shown in Supplementary Table S1. The tumors used for GIScar validation are shown in Supplementary Table S2. The source code of GIScar as well as the protocol for the library preparation are available on request. The sequencing data from the tumor collections cannot be provided in a public repository; the sequencing data from tumors of the patients included in the PAOLA-1 clinical trial are the propriety of the ARCAGY-GINECO group. Therefore, the sequencing data are available on reasonable request.

HRD status was available by MGMC for 210 of the 250 DNA tumor samples in the prospective collection. Data were missing from the remaining 40 samples because they: (i) did not meet the MGMC criteria for analysis or MGMC could not compute a score (24/40 samples; 60%); (ii) were not sent for MGMC (9/40; 23%); and (iii) were sequenced before the French national authorization of olaparib treatment for HRD-positive ovarian cancer (thus not sent for MGMC, 7/40; 18%). MGMC HRD-positive tumors represented 40% of ovarian tumors (100/250) and among these, 36% (36/100) were BRCA mutated (Fig. 1D).

Among the 469 DNA samples from the PAOLA-1 collection, HRD status could not be established by MGMC for 44 tumors. The proportion of HRD-positive tumors (242/469; 52%) was higher than in the prospective collection (39%). This difference was explained mainly by an increase in BRCA-mutated tumors [32% (150/469) vs. 14% (36/250); Fig. 1E].

GIScar was built on sequencing data from the prospective collection. The purity recalculation by the ASCAT method revealed a greater degree of variation than initially estimated by the pathologists (Supplementary Fig. S1A). Consequently, we considered only the ASCAT purity to compute data by the CNVkit pipeline. Sequencing of the 127-gene panel revealed a typical value of 400 SNP detected per sample (Supplementary Fig. S1B). This number of SNP allowed CNVkit estimation of B-allele frequency for most (75.9%) genomic regions (data not shown). In the prospective collection of 250 tumors, the average sequencing depth on targeted regions was 1,579×, ranging from 16× to 3,300×. In the clinical collection of 469 tumors, the average sequencing depth was 653×, with a range of 0× to 1755×. Scores for LGE, SIS, and AI calculated from the CNVkit pipeline output showed similar Pearson correlation values, ranging from 0.78 to 0.81 (Supplementary Fig. S1C). Correlation values were reduced when SIS and AI scores were normalized by chromosome size.

GIScar score training after variable selection revealed that the scores for LGE, SIS, and AI played a significant role in predicting HRD status (LGE, P_{Wald test} = 4.27× 10⁻¹⁵; SIS, P_{Wald test} = 1.97× 10⁻⁷; AI, P_{Wald test} = 0.00406; AInorm, P_{Wald test} = 3.56 × 10^-5). The GIScar model was improved only by normalizing the AI score.

When applied to the prospective collection, GIScar detected 109 (44%) of 250 HRD-positive tumors (Fig. 2A). The same BRCA pathogenic variants were detected in samples sequenced by MGMC and by our platform (Supplementary Table S5). The ROC analysis indicated that combining the three scores (LGE, SIS, and AI) improved concordance with MGMC classification compared with use of the individual scores (Fig. 2B). The AUC was 0.935 for GIScar and 0.888, 0.734, and 0.720, respectively, for LGE, AI, and SIS. The optimal threshold observed from ROC data was 0.48 (i.e., a tumor with a GIScar score ≥ 0.48 was considered as HRD-positive). At this threshold, the sensitivity was 88.0% (88/100) and the specificity was 90.0% (99/110). The percent of agreement was 89% (187/210) and 87% (151/174) when excluding the BRCA-mutated tumors. The two scores (MGMC and GIScar) showed significant correlation at the HRD classification level (χ² test P < 0.001). The kappa statistics were 0.780 and 0.715 if the BRCA-mutated tumors were excluded, representing close agreement. The 12 false negatives in the prospective collection had a median MGMC score of 46 (upper limit 59), whereas the 88 true positives had a median MGMC score of 69 (upper limit 94). These data suggest that the GIScar false-negative tumors could have a lower level of instability than the true-positive cohort. The score values of MGMC and GIScar showed a significant correlation with the Pearson coefficient of 0.792 (P < 0.001; Fig. 2C). The distribution of GIScar scores (Fig. 2D) revealed that 85% (31/36) of BRCA-mutated tumors had a GIScar score above the threshold of 0.48. Among the five remaining BRCA-mutated tumors, two tumors had positive scores by MGMC test (70 and 76) and three had negative/missing scores. These results confirmed the ability of the GIScar test to detect HRD-positive tumors.

The HRD-positive tumors detected by GIScar that also harbored BRCA mutations had significantly better PFS on olaparib treatment compared with placebo. Median PFS (mPFS) was 42.6 months with olaparib and 19.5 months with placebo [HR, 0.45; 95% confidence interval (CI), 0.33–0.62; log-rank P < 0.00001; Fig. 3A]. The HRD-positive tumors without BRCA mutation also had significantly better PFS with olaparib than placebo, with mPFS of 27.6 versus 16.3 months, respectively (HR, 0.50; 95% CI, 0.31–0.80; log-rank P = 0.0030; Fig. 3B). Conversely, HRD-negative tumors did not derive a significant benefit from olaparib, with mPFS of 16.6 months in the olaparib arm versus 16.6 months in the placebo arm (HR, 1.02; 95% CI, 0.74–1.40; log-rank P = 0.90; Fig. 3C).

When comparing PFS outcomes by treatment arm according to HRD status, we observed a significant PFS difference according to HRD status in the olaparib arm (mPFS: 16.6 vs. 42.6 months in HRD-negative vs. HRD-positive tumors, respectively; HR, 0.34; 95% CI, 0.26–0.44; log-rank P < 0.00001). However, we observed no association between PFS and HRD status in the placebo arm (mPFS: 16.6 vs. 19.5 months in HRD-negative vs. HRD-positive tumors; HR, 0.77; 95% CI, 0.54–1.11; log-rank P = 0.16; Supplementary Fig. S2).

Assessment of HRD status according to MGMC and GIScar test on the clinical collection revealed that the number of HRD-positive tumors detected was higher with GIScar than MGMC (258 vs. 242 tumors, respectively; Fig. 4A). This difference was partially explained by the lower proportion with inconclusive results by GIScar [4/469 (1%) versus 44/469 (9%) with MGMC]. The correlation between MGMC and GIScar scores was lower than for the prospective collection (Fig. 4B). Excluding the missing data, the percent of agreement was 89% (377/422) and 84% (230/275) excluding the BRCA-mutated tumors. The two scores (MGMC and GIScar) showed significant correlation at the HRD classification level (χ² test P < 0.001). The kappa statistics were 0.781 overall and 0.643 when excluding the BRCA-mutated tumors, representing close agreement. Interestingly, the GIScar score in discordant cases tended to be closer to the decision-making threshold than the score in concordant cases. The average score of tumors with an HRD-negative status by GIScar tests was 0.15 in samples with HRD-negative MGMC status and 0.32 in samples with HRD-positive MGMC status. The average score of tumors with an HRD-positive status by GIScar was 0.70 in samples with HRD-positive MGMC status and 0.60 in samples with HRD-negative MGMC status (Supplementary Table S6).

Among BRCA-wild-type tumors with matched samples (i.e., excluding those with missing data for either MGMC or GIScar), both tests showed a significant benefit from olaparib for HRD tumors and not for HRD-negative tumors (Fig. 4C). We observed no significant difference between the two tests. The 95% CI for the HR were largely overlapping for HRD-positive tumors. For samples that were HRD-positive by MGMC, mPFS was 17.6 versus 38.9 months with placebo versus olaparib, respectively (HR, 0.43; 95% CI, 0.26–0.72; log-rank P = 0.00083) and for samples that were HRD-positive by GIScar, mPFS was 16.1 versus 23.9 months, respectively (HR, 0.49; 95% CI, 0.30–0.81; log-rank P = 0.0037). Among HRD-negative tumors, those identified as HRD-negative by MGMC had mPFS of 16.5 versus 16.6 months with placebo versus olaparib, respectively (HR, 1.07; 95% CI, 0.76–1.52; log-rank P = 0.70), and those identified as HRD-negative by GIScar had mPFS of 17.0 versus 16.8 months with placebo versus olaparib, respectively (HR, 1.08; 95% CI, 0.76–1.53, log-rank P = 0.66; Fig. 4C). Among samples with discordant GIScar and MGMC classification, tumors identified as HRD-positive by MGMC but HRD-negative by GIScar showed a trend toward better PFS on olaparib, but in this relatively small subgroup, the difference in mPFS (33.3 months with olaparib vs. 27.7 months with placebo; HR 0.39, 95% CI, 0.12–1.26) did not reach statistical significance (log-rank P = 0.10; Fig. 4D). On the other hand, tumors that were HRD-positive by GIScar but HRD-negative by MGMC had significantly better PFS with olaparib, with mPFS of 16.8 months with olaparib versus 9.7 months with placebo (HR, 0.23; 95% CI, 0.07–0.72; log-rank P = 0.006; Fig. 4D). Because of the limited number of patients in this subgroup (n = 27), we were unable to prove the superiority of GIScar over MGMC. Tumor HRD classification by GIScar in the subset of samples with inconclusive HRD status by MGMC correlated significantly with PFS, with mPFS exceeding 40 months in the olaparib arm versus 15.4 months in the placebo arm (HR, 0.30; 95% CI, 0.10–0.85; log-rank P = 0.016; Supplementary Fig. S3). As only nine samples were identified as HRD-positive by GIScar but inconclusive by MGMC, it was not possible to assess the impact of olaparib in this subgroup. However, none of the three patients with HRD-positive tumors by GIScar who received olaparib had disease progression during follow-up (Fig. 4E). Conversely, patients with HRD-negative tumors by GIScar and inconclusive by MGMC showed no relevant benefit from olaparib (mPFS 15.6 months with olaparib vs. 11.0 months with placebo; HR, 0.74; 95% CI, 0.35–1.57; log-rank P = 0.44; Fig. 4E).

Assessment of the correlation between score value and PFS indicated that tumors identified as HRD-negative by both GIScar and MGMC showed similar PFS regardless of the score value (Supplementary Fig. S4A). PFS in the placebo group was similar to PFS in patients with a negative score treated with olaparib (Supplementary Fig. S4B). Tumors with a positive score for MGMC showed highly variable PFS, which did not appear to be correlated with the score value. For example, the tumors with a score ranging from 44.5 to 56 had longer PFS than tumors with a score ranging from 56 to 63.2 (Supplementary Fig. S4A). By contrast, tumors with a positive score by GIScar showed homogeneous PFS.

The performance of GIScar was confirmed for OS. Among patients whose tumors were HRD-positive by GIScar, OS was significantly better in the olaparib than the placebo group (median OS 75.2 vs. 59.8 months, respectively; HR, 0.66; 95% CI, 0.45–0.98; log-rank P = 0.037; Fig. 5A). In contrast, patients with HRD-negative tumors by GIScar derived no benefit from olaparib. In this subgroup, olaparib maintenance therapy showed numerically worse outcomes compared with placebo, with a median OS of 39.3 months versus 46.3 months, respectively (HR, 1.24; 95% CI, 0.86–1.78; log-rank P = 0.25; Fig. 5B). Comparing outcomes on olaparib therapy according to GIScar result, median OS was 39.3 months in HRD-negative tumors versus 75.2 months for HRD-positive tumors.

In these analyses, we show that the GIScar test, which was developed on the basis of the unique sequencing of a gene panel to detect both BRCA variants and genomic instability, reliably detects HRD-positivity and predicts sensitivity to olaparib with bevacizumab maintenance therapy in patients with ovarian cancers. GIScar showed high analytical concordance with the MGMC test. The training set of 250 ovarian cancers showed high correlation (ROC_AUC 0.935) with the historical commercial test. PFS and OS HRs with GIScar were similar to the performance of the MGMC test, indicating the ability to identify patients deriving benefit from olaparib.

From a practical perspective, the GIScar test reduced the proportion of inconclusive cases by 90% compared with the MGMC test. GIScar-defined HRD status correlated with sensitivity to olaparib therapy, even in samples for which HRD status could not be determined using the MGMC test. We observed that in the clinical collection, cases discordant between GIScar and MGMC, tended to have a GIScar score close to the decisional threshold compared with concordant cases (Supplementary table S6). We hypothesize that these tumors with discordant status have an intermediate level of genomic instability. The tumor heterogeneity of the FFPE tissue should also be taken into account, as MGMC and GIScar tests were not performed on the same DNA extraction. In addition, the FFPE DNA samples from PAOLA-1 were collected at the time of patient enrollment (between 2015 and 2016) but not sequenced in this study until 2022. Therefore, some of the DNA samples could be partially decayed compared with freshly extracted FFPE DNA samples.

Collaboration with other centers in France demonstrated that GIScar can be easily implemented into other diagnostic laboratories, with important implications for selecting patients likely to benefit from maintenance therapy with a PARP inhibitor. The library preparation requires the same material with a similar workflow to the sequencing of BRCA1/2 genes alone. The source code to compute GIScar score could be implemented in an automatic pipeline if the laboratory has its own bioinformatics resources. The source code of GIScar as well as the protocol for the library preparation are available on request. Furthermore, the turnaround time for GIScar takes less than 5 days once tumor DNA is available, making it highly suitable for clinical diagnosis.

Numerous academic as well as commercial HRD tests have been developed in Europe. They provide the opportunity of sharing the HRD test with any genomic laboratory having sufficient NGS capacity. Most of them have shown good performance in identifying HRD-positive tumors compared with MGMC within the ENGOT HRD European Initiative (unpublished data). Although there are several differences between the various tests that have been validated using samples from PAOLA-1, most offer an important benefit in terms of success rate, substantially reducing the test failure rate. In addition, all have shown good correlation with MGMC and can predict PFS and OS with maintenance olaparib in newly diagnosed ovarian cancer. Furthermore, they may be considerably less expensive, and it is encouraging that available alternatives can be used with confidence. The MGMC test costs more than $4,000 per analysis (29), whereas the average cost for the GIScar test is $750 per analysis, of which $240 is the cost of wet lab reagents. Of note, the turnaround time between receipt of the FFPE sample and the final result of MGMC and GIScar is similar, with an average turnaround time of 2 weeks.

In addition, our method requires sequencing of only a limited panel of genes (size of the capture set: 650 Kb), whereas the other HRD tests are based on the sequencing of whole-genome (WGS) or large panels including thousands of polymorphisms. Consequently, our method decreases the sequencing capacity and thus the sequencing cost. As an example, it is possible to sequence up to 48 samples per 2×75 high-output flow cell on a NextSeq 500 Illumina sequencer. Importantly, the test failure rate remained particularly low (1%), probably because the regions sequenced are smaller than with the other HRD tests.

Regarding the bevacizumab impact, we observed no significant difference in PFS between HRD-positive tumors and HRD-negative tumors. We hypothesized that HRD status did not impact the efficiency of the antiangiogenic drugs. However, the lack of difference could be explained by the small sample size. The potential interest of other PARP inhibitors and monotherapy may be explored in the future. These perspectives underline the importance of sharing tumor samples or clinical data from pharmaceutical industry.

Although we consider our results to be extremely encouraging, we acknowledge potential criticisms. Firstly, despite GIScar significantly reducing the failure rate compared with MGMC, it was not superior to MGMC regarding predictive value for PFS. Concordance with MGMC was high, but we recognize that concordance with an imperfect gold standard leaves room for improvement (21, 30, 31). The limitation of testing approaches that rely on sequencing a panel of genes known to cause HRD is that the relationship between HRD and PARP inhibitor sensitivity is imperfect (31). Some non-BRCA HRR genes do not predict response to PARP inhibition, and additional HRR gene alterations not included in current tests may be associated with sensitivity (30). In a recent analysis of the PAOLA-1 trial, non-BRCA HRR mutation gene panels were not predictive of PFS benefit from olaparib (32). Among the 496 samples tested with GIScar in our study, 41 tumors carried pathogenic or likely pathogenic variants in HRR genes (excluding BRCA1/2 genes). Nevertheless, patients with mutations in PALB2 (n = 2), RAD51C (n = 6), and RAD51D (n = 6) seemed to show a longer PFS, but the number of patients was insufficient to perform appropriate statistical analyses. To some extent, this is overcome by approaches that detect genomic scars (chromosomal aberrations resulting from HRR pathway dysfunction), as with MGMC. MGMC combines both strategies to test HRR mutations and HRD genomic instability. However, tumors that are no longer “functionally” HRD because of undetected reversion mechanisms may still be identified as HRD-positive. An emerging investigational approach that avoids this difficulty is to test functional HRD, determining real-time homologous recombination status, for example by testing the ability of RAD51 protein to recruit nuclear RAD51 foci (33, 34). However, these strategies have not yet been validated. There is currently insufficient evidence that they predict sensitivity to PARP inhibitors, and they are unsuitable for use in routine clinical practice (22, 35).

Another limitation of the present analyses is that we validated GIScar using a retrospective sample collection. In addition, the PAOLA-1 trial was not stratified based on HRD status. However, testing the different academically developed assays on an identical set of trial samples should enable elucidation of the strengths and limitations of each assay. The next step will be to validate GIScar using a prospective dataset and also to test the ability of the assay to predict sensitivity to other PARP inhibitors in routine practice (niraparib and rucaparib) or in development.

In summary, we believe GIScar is a valuable diagnostic tool with potential to improve patient selection for PARP inhibitor therapy and ease the distribution of testing among genetics laboratories. Exploration and development in other tumor types is also warranted, given the broader application of PARP inhibition to other solid tumor types, including breast, pancreatic, and prostate cancers.

F. Joly reports personal fees from AZ-MSD and GSK outside the submitted work and reports other support (travels for congress) from MSD. L.-M. Chevalier reports personal fees from AstraZeneca, MSD, and SophiaGenetics outside the submitted work. E. Rouleau reports grants from AstraZeneca, Clovis, BMS, GSK, Roche Diagnostic, and MSD during the conduct of the study. A. González-Martín reports personal fees from Alkermes, Amgen, AstraZeneca, Clovis Oncology, Genmab, GSK, ImmunoGen, Merck Sharp & Dohme, Novartis, Oncoinvent, Pfizer/Merck, PharmaMar, Roche, Sotio, and Sutro outside the submitted work. H.-J. Lück reports personal fees from AstraZeneca, Gilead, Lilly, and GSK outside the submitted work. I. Ray-Coquard reports personal fees from CLOVIS, AstraZeneca, and Roche and grants and personal fees from GSK outside the submitted work. E. Pujade-Lauraine reports grants from AstraZeneca during the conduct of the study; personal fees from AstraZeneca/Merck, Agenus, and Incyte; personal fees and nonfinancial support from GSK; and other support from ARCAGY Research outside the submitted work. D. Vaur reports grants from AstraZeneca and personal fees from GSK outside the submitted work. No disclosures were reported by the other authors.

R. Leman: Conceptualization, data curation, supervision, validation, visualization, methodology, writing–original draft, project administration, writing–review and editing. E. Muller: Conceptualization. A. Legros: Resources. N. Goardon: Resources. I. Chentli: Resources. A. Atkinson: Resources. A. Tranchant: Resources. L. Castera: Resources, writing–review and editing. S. Krieger: Resources, writing–review and editing. A. Ricou: Resources, writing–review and editing. F. Boulouard: Resources, writing–review and editing. F. Joly: Resources, writing–review and editing. R. Boucly: Resources, writing–review and editing. A. Dumont: Resources, writing–review and editing. N. Basset: Resources, writing–review and editing. F. Coulet: Resources, writing–review and editing. L.-M. Chevalier: Resources, writing–review and editing. E. Rouleau: Resources, writing–review and editing. K. Leitner: Resources, investigation. A. González-Martín: Resources, investigation. P. Gargiulo: Resources, investigation. H.-J. Lück: Resources, investigation. C. Genestie: Resources, investigation. PAOLA-1 Investigators: Resources. I. Ray-Coquard: Resources, data curation, investigation, writing–review and editing. E. Pujade-Lauraine: Resources, data curation, formal analysis, writing–review and editing. D. Vaur: Validation, project administration, writing–review and editing.

We thank Christine Montotot and ARCAGY team for organization assistance. We thank the Biological Resource Center (BRC) of the Institut Curie for supplying PAOLA-1 DNA biobank material. We thank the GENECAN platform for laboratory assistance. We would like to acknowledge the SéSAME (Séquençage pour la Santé, l'Agronomie, la Mer et l'Environnement) platform for sequencing the DNA samples. We thank the medical writer Jennifer Kelly of Medi-Kelsey Ltd for her writing assistance, technical editing, language editing, and proofreading.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.