We analyzed genomic instability data from six phase II or III studies evaluating olaparib as maintenance therapy (SOLO1 [8, 13], PAOLA-1 [9, 22], Study 19 [14,15,16], SOLO2 [18, 21], OPINION [19, 20]) or as treatment (LIGHT [12, 17]). Our study comprehensively describes the relationship between GIS and a 108-gene panel in > 2000 patients with newly diagnosed [8, 9, 13, 22] or platinum-sensitive relapsed [12, 14,15,16,17,18,19,20,21] high-grade serous or high-grade endometrioid ovarian cancer, and to our knowledge represents the largest combined HRD analysis of high-grade epithelial ovarian cancer using a commercially available HRD assay. To our knowledge, for the first time, our study shows the relationship between GIS and a broad panel of genes in carefully curated patients without BRCAm or HRRm.

We demonstrated that GIS has an overall bimodal distribution in high-grade serous or high-grade endometrioid ovarian cancer, with high GIS in BRCAm tumors, variable GIS in non-BRCA HRRm tumors (possibly reflecting differences in the role of these genes in the HRR pathway), and predominantly low GIS in non-HRRm tumors. This pattern was consistent across individual studies and in patients with newly diagnosed ovarian cancer or PSROC.

In our analysis and consistent with the individual studies [7, 25, 26], BRCAm was associated with high GIS and high rates of biallelic loss, irrespective of BRCAm being germline or somatic in origin. Our analysis also demonstrated that GIS was high regardless of mutation subtype. Although both BRCA1m and BRCA2m tumors had a very high GIS (with a similar IQR), a significantly higher score was observed for those with BRCA1m. Clinical trial data have demonstrated significant benefit of PARP inhibitor treatment in ovarian cancer patients with BRCAm; however, some studies suggest that patients with BRCA2m have relatively greater sensitivity to PARP inhibition than those with BRCA1m [27, 28]. In the SOLO1 trial, benefit from olaparib was demonstrated in patients with a BRCAm regardless of whether genome-wide LOH scores were high or low [25]. These apparently discordant observations may reflect the fact that GIS is only a surrogate marker for sensitivity to PARP inhibition and not the sole determinant. It is possible that once the threshold for GIS has been passed, other factors, such as drug resistance mechanisms, play a more prominent role in tumor sensitivity to PARP inhibition. Further work is needed to understand the relationship between BRCAm and GIS in BRCAm tumors with low genomic instability.

As shown in phase III trials [9,10,11], HRD, as determined by genomic instability testing, is critical in the newly diagnosed ovarian cancer setting to identify which patients may experience the greatest benefit from PARP inhibitor maintenance therapy [29, 30]. HRD is a measure of global genomic instability induced by defects in the HRR pathway, while HRRm reflects the presence of mutations in specific genes involved in the HRR pathway. Our data demonstrate that HRD and HRRm analyses are not interchangeable and identify different sub-populations of patients. The finding that there was no statistically significant difference in median GIS across the non-BRCA HRRm population in the newly diagnosed ovarian cancer and PSROC settings was unexpected as it suggests that a difference in GIS does not explain why HRRm was predictive of PARP inhibitor benefit in Study 19 [7], but not in PAOLA-1 [31]. Some selection for patients with an overall higher GIS might have been expected in the platinum-sensitive relapsed setting given these studies only included patients known to have platinum-sensitive disease, whereas patients with no evidence of disease following cytoreductive surgery and whose sensitivity to platinum was unknown were eligible for inclusion in the first-line studies. Non-BRCA HRRm were low in prevalence, not enriched in HRD-positive tumors, and heterogeneous with regards to biallelic loss and GIS. RAD51C, RAD51D, BRIP1, and PALB2 are genes known to be associated with a hereditary risk of ovarian cancer [32,33,34]. The small number of PALB2 mutations limits interpretation; however, RAD51C, RAD51D, and BRIP1 mutations were associated with both high rates of biallelic loss and high GIS, suggesting that these genes might be true drivers of HRD in ovarian cancer. Mutations in these genes accounted for the majority of non-BRCA HRRm in OPINION, compared with PAOLA-1, LIGHT, and Study 19 where other non-BRCA HRRm genes were also prevalent. This could explain why we observed a higher degree of GIS in OPINION relative to the other studies. In terms of other non-BRCA HRR genes, studies suggest that loss-of-function mutations in CDK12 confer sensitivity to PARP inhibition [35]. A possible explanation for the relatively low GIS associated with CDK12 in this analysis is that CDK12 mutations have a distinct genomic instability pattern characterized by focal tandem repeats [36] and the Myriad MyChoice® CDx assay does not detect this particular genomic instability signature. Our knowledge of what constitutes an HRRm gene continues to evolve. Only two patients with variants in PPP2R2A were included in this analysis; low GIS was observed in both cases. After the studies included in this analysis were initiated, data from the PROfound study demonstrated that no benefit of olaparib over control therapy was observed in patients with prostate cancer and PPP2R2A mutations and was not predictive of benefit from PARP inhibitor therapy [37].

In terms of alterations outside of BRCAm or HRRm, those in MAP3K1, MYC, RB1, KMT2D, CSMD3, SOX2, NF1, MAP2K4, and NTHL1 had a high median GIS and those in PTEN, STAG2, PTPRD, FANCM, ERBB2, TSC1, MCL1, BLM, PIK3CA, MYH, NBN, CCNE1, ARID1A, NRAS, and KRAS had a low median GIS. It is now understood that RB1 loss is not mutually exclusive with HRRm, and improved outcomes have been reported when RB1 loss co-occurs with HRRm [38, 39]. This analysis demonstrates that even in tumors without BRCAm or HRRm, NF1 and RB1 mutations were associated with relatively high genomic instability; further work is needed to understand the mechanistic relationship between NF1 and RB1 and GIS. By contrast, mutations in CCNE1 and PIK3CA were associated with relatively low genomic instability that was found to be significantly lower than that seen in non-CCNE1-altered and non-PIK3CA-mutated tumors, respectively, in non-BRCAm states. To our knowledge, PIK3CA mutations have not previously been shown to be associated with low GIS and an association between CCNE1 amplification and low GIS has previously only been observed in smaller datasets [40]. Increased cyclin E expression, either by CCNE1 gene amplification, copy-number gain, or elevated protein expression, is associated with poor clinical outcomes and resistance to DNA-damaging drugs in ovarian cancer [24, 41]. For example, CCNE1 amplification was correlated with shorter relapse-free survival in patients with ovarian carcinomas treated with platinum-based chemotherapy [41]. The results of our analysis and others [42, 43] suggest that patients with CCNE1 amplifications (but without BRCAm) may potentially benefit from alternative targeted therapies. PIK3CA mutations are also associated with poor response to platinum-based chemotherapy and platinum resistance in patients with ovarian cancer [44]. Patients with alterations in CCNE1 or PIK3CA therefore represent a population with very high unmet medical needs. The optimal treatment of CCNE1-amplified and PIK3CA-mutated tumors warrants further investigation; the PI3K/AKT/mTOR pathway is one of the potential therapeutic targets for CCNE1-mutated [45] and PIK3CA-mutated [44] tumors.

Ovarian cancer is the archetypal homologous recombination deficient tumor and is the only tumor type to date where genomic instability status is predictive of benefit from PARP inhibitor therapy. Although correlating genomic instability with clinical benefit was beyond the scope of the current analysis, previous analyses of PAOLA-1 [9, 22], Study 19 [7], OPINION [19, 20], and LIGHT [12, 17] have evaluated clinical outcome according to HRD status. Further work is needed to understand the clinical implications of these data and to evaluate potential links between genomic instability and clinical benefit from PARP inhibitor therapy for individual non-BRCA HRR and non-HRR genes. To date, post hoc exploratory analysis found that non-BRCA HRR gene panels were not predictive of the efficacy of maintenance olaparib plus bevacizumab in PAOLA-1 [31]. By contrast, subgroup analyses based on non-BRCA HRRm demonstrated the benefit of olaparib treatment in Study 19 [7], olaparib activity comparable with that seen in BRCAm in the ORZORA trial [46], and longer PFS benefit in patients with non-BRCA HRRm relative to patients with non-HRRm in the OPINION trial [47]. One explanation for these apparent discrepancies might be a difference in signal between the first-line [31] and relapsed disease [7] settings, with the relapsed disease population more likely to be enriched for platinum sensitivity. These studies included small numbers of individual non-BRCA HRR genes and differences between the studies in gene-to-gene prevalence may be another potential explanation for apparent discrepancies. For example, of the non-BRCA HRRm subgroup, CDK12, BRIP1, and RAD51C accounted for 24%, 13%, and 17%, respectively, in PAOLA-1 [31]; 36%, 15%, and 18%, respectively, in ORZORA [46]; and 9%, 21%, and 24%, respectively, in OPINION [47]. Furthermore, these studies were not individually powered to assess the predictive power of these non-BRCA HRR genes, making comparisons difficult. In terms of other PARP inhibitors, RAD51C and RAD51D mutations predicted response to treatment with rucaparib patients with relapsed ovarian cancer in a post hoc exploratory analysis of ARIEL2 [48]. Although we found BRIP1 mutations had a high GIS, they are yet to be established as predictive of PARP inhibitor response [31]. Better understanding of drivers of genomic instability in ovarian cancer may open up opportunities for PARP inhibitor use in other indications where GIS alone has not been shown to be predictive. Our study identified a further nine genes (MAP3K1, MYC, RB1, KMT2D, CSMD3, SOX2, NF1, MAP2K4, and NTHL1) which, when mutated, were associated with high median GIS and an additional 15 genes (PTEN, STAG2, PTPRD, FANCM, ERBB2, TSC1, MCL1, BLM, PIK3CA, MYH, NBN, CCNE1, ARID1A, NRAS, and KRAS) which, when mutated, were associated with low median GIS; some of these genes (e.g., CCNE1, ARID1A, NRAS, and KRAS) are linked with replication stress. Thus, our data highlight the heterogeneity of HRD-positive and HRD-negative populations, which appear to have different biology, and the need to better personalize treatment options that target alternative pathways. Further clinical studies would be required to validate these targets as biomarkers for PARP inhibitor benefit in ovarian cancer, which is beyond the scope of this study. These data also signal the need to explore PARP inhibitor benefit in other tumor types harboring mutations shown here to associate with high GIS and, potentially, an HRD phenotype.

This analysis is associated with several limitations. The first is pooling data from six different trials—with variations in study design, treatment setting, and patient selection criteria as well as diagnostic tests applied—which may have contributed to differences in the results between studies. For example, these data do not represent the patterns of genomic instability in an all-comer population, as some of the trials selected for biomarkers prospectively (OPINION [19, 20], SOLO1 [8, 13], SOLO2 [18, 21], LIGHT [12, 17]), whereas others did not (PAOLA-1 [9, 22], Study 19 [14,15,16]). Hence, the prevalence of BRCAm, HRRm, and non-HRRm is not reflective of all patients with ovarian cancer. In addition, LIGHT was conducted in the later-line treatment setting in patients who had received one or more prior lines of platinum-based chemotherapy [12, 17], whereas the other studies evaluated olaparib maintenance therapy in patients who responded to platinum-based chemotherapy (in combination with bevacizumab in PAOLA-1 [9, 22]). It is not clear to what extent this heterogeneity may have impacted the pooled dataset used in this analysis. Secondly, while BRCA1 methylation data are available from Study 19 [7] and PAOLA-1 [49], methylation data were not available from the other studies included in this analysis. Hypermethylation may partly explain the genomic instability seen in some non-HRRm tumors. BRCA1 hypermethylation is found in approximately 10% of ovarian tumors [50] and is associated with high levels of genomic instability [7]. In PAOLA-1, of the 72 BRCA1 or RAD51C methylated tumors with a valid GIS, 92% had a GIS ≥ 42 [49]. Similarly, a recent study in high-grade serous ovarian carcinoma showed that homozygous methylation of the RAD51C promoter is predictive of sensitivity to PARP inhibition [51]. Lack of methylation data was a limitation when assessing genomic instability in the non-BRCAm non-HRRm subgroup. A further limitation is the possibility that some large rearrangements in BRCA (germline or somatic) and some CCNE1 amplifications may not have been detected because of the assays used, causing tumors with these alterations to be underrepresented in the analysis. In addition, data regarding whole-genome duplication were not available for this analysis; however, minor allele frequency was assessed to ensure it was at zero when evaluating biallelic loss. Finally, this analysis only includes clinical trials of olaparib and may not be representative of all PARP inhibitors.

In summary, this analysis of genomic data from six olaparib studies [