Preclinical In Vivo Validation of the RAD51 Test for Identification of Homologous Recombination-Deficient Tumors and Patient Stratification

PARP inhibitors (PARPi) have been approved for the treatment of ovarian (HGSOC), breast, prostate, and pancreatic cancers (PaC; ref. 1). In these settings, biomarkers such as sensitivity to platinum-based chemotherapy and homologous recombination repair (HRR) deficiency (HRD) tests enable patient selection for these targeted therapies. Among HRD tests, pathogenic variants in genes in the HRR pathway or genomic HRD tests—scars or signatures—are the most commonly used (2–4).

Overall, it is estimated that over 17% of all cancers harbor genetic alterations in the HRR pathway (5), albeit only those that are biallelic lead to HRD (2, 3). Genomic scars or signatures have shown sensitivity in identifying tumors with an HRD phenotype (4, 6, 7). In patients with HGSOC, current genomic scars are informative for predicting the magnitude of benefit from PARPi as maintenance therapy after response to platinum-based chemotherapy (8). Nevertheless, genomic HRD tests may fail to provide an accurate functional status of HRR or predict the benefit from PARPi and other DNA damaging agents in other settings. Also, these assays carry some challenges for their implementation in clinical practice, because they are technically complex, may require frozen samples, and carry a high cost. In this regard, some commercially-available panels based on targeted next-generation sequencing (NGS) have incorporated measures of HRD based on surrogate genomic scars (7). Although the cost of targeted NGS testing is significantly less than for whole-exome sequencing (WES), it still represents a barrier for many patients and healthcare systems. In summary, genomic testing presents technical and logistical challenges to reach large numbers of patients with cancer. In comparison, histology-based tests are technically less complex than NGS tests, require less amount of sample, are less costly and can capture tumor heterogeneity and changes in HRR status during tumor evolution. As such, previous evidence has shown that detection of the downstream HRR protein RAD51 nuclear foci is a marker of HRR functionality (9–11).

We and others have adopted preclinical models that recapitulate the patient's disease and help investigate biomarkers of sensitivity to targeted therapies, including cell lines (12), patient-derived xenografts (PDX) models (13–15), and three-dimensional patient-derived tumor cell cultures (PDC; refs. 16–19). Using PDX, we developed an immunofluorescence-based test to detect RAD51 nuclear foci in formalin-fixed paraffin-embedded (FFPE) samples (14, 15). The clinical utility of the assay has been tested in early triple-negative breast cancer (TNBC) and in castration-resistant prostate cancer treated with the PARPi olaparib (20–22). This RAD51 assay is able to capture: (i) HRD in tumors with BRCA1, BRCA2, or PALB2 mutations, or in tumors with epigenetic silencing of an HRR gene, that is, in BRCA1 or RAD51C; or (ii) HRR proficiency (HRP) in tumors with previous history of HRD that have restored HRR as a mechanism of drug resistance (14, 15).

In summary, there is a clear need to provide access to HRD functional tests available for routine clinical diagnostics to identify patients eligible to targeted therapies. In this work, we tested the accuracy of the RAD51 test in comparison with mutations in HRR genes or genomic HRD tests to predict response to PARPi, with further validation in clinical samples. We also conducted an analysis of the predictive value of the RAD51 assay for predicting response to platinum. The overarching aim was to provide preclinical analytical validation of the RAD51 assay to support the use of the test in clinical practice. In addition, we described potential mechanisms of resistance to PARPi, thus cataloguing the PDX models for further experimental studies.

Fresh tumor samples from patients with HRR-related breast (triple negative, luminal B, or HER2-positive), HGSOC or pancreatic cancer were collected for diagnostic and for implantation into nude mice under an Institutional Review Board (IRB)–approved protocol. The human biological samples were sourced ethically and their research use was in accord with the terms of the written informed consents under an IRB/EC approved protocol. Experiments were approved by the Ethical Committee of Animal Experimentation of the Vall d'Hebron Research Institute. All animal studies were ethically reviewed and carried out in accordance with European Directive 2010/63/EEC and the GSK Policy on the Care, Welfare and Treatment of Animals. The PARPi response criteria was based on the percentage of tumor volume change. Responders included models that exhibited complete response (CR), partial response (PR), and stable disease (SD) and nonresponders included progressive disease (PD). For further details, see Supplementary Materials and Methods.

DNA was extracted from PDX samples using DNeasy Blood & Tissue Kit (Qiagen). For further details, see Supplementary Materials and Methods. Three genomic scar/signature assays were performed: the Myriad's myChoice® HRD test carried out at Myriad Genetics on DNA extracted from a subset of Xentech's PDX (using NucleoBond AXG100 Kit, Macherey- Nagel), HRDetect and the genomic instability score (GIS: levels of allelic imbalance, loss of heterozygosity, number of large-scale transitions) performed at Wellcome Trust Sanger Institute on DNA extracted from a subset of VHIO's PDX as described in refs. 6 and 23, respectively.

BRCA1 promoter methylation was measured using methylation-specific multiplex ligation-dependent probe amplification (MS-MLPA; MRC Holland) according to manufacturer's instructions. Two of the xenografts generated in CRUK/UCAM (STG139 and STG201) had been previously tested using reduced-representation bisulfite sequencing (RRBS; ref. 16) and further validated using MS-MLPA. Positive controls of BRCA1 promoter hypermethylation were used (T127 and/162; ref. 24).

RNA was extracted from PDX samples (15–30 mg) by using the PerfectPure RNA Tissue Kit (five Prime). The purity and integrity were assessed by the Agilent 2100 Bioanalyzer system, and RNA-seq was performed in Illumina HiSeq 4000 (Illumina). Log2 transformation was used for data analysis.

The following primary antibodies were used for immunofluorescence: rabbit anti-RAD51 (Abcam ab133534, 1:1,000), mouse anti-geminin (NovoCastra NCL-L, 1:100 in PDX samples, 1:60 in patient samples), rabbit anti-geminin (ProteinTech 10802–1-AP, 1:400), mouse anti-BRCA1 (Santa Cruz Biotechnology, sc-6954, 1:50), mouse anti-BRCA1 (Abcam, ab16780, 1:200), mouse anti-phospho-H2AX (Millipore, #05-636, 1:200). Goat anti-rabbit Alexa Fluor-568 (Invitrogen; 1:500), goat anti-mouse Alexa Fluor-488 (Invitrogen; 1:500), donkey anti-mouse Alexa Fluor-568 (Invitrogen; 1:500), and goat anti-rabbit Alexa Fluor-488 (Invitrogen; 1:500) were used as secondary antibodies. The immunofluorescence was performed as described in ref. 15.

Biomarkers were quantified on FFPE PDX, PDC, or patient tumor samples by scoring the percentage of geminin-positive cells with five or more nuclear foci of any size (15). Geminin is a master regulator of cell-cycle progression that enables to mark for S/G2-cell cycle phase (25). Scoring was performed onto life images using a 60x-immersion oil lens. One hundred geminin-positive cells from at least three representative areas of each sample were analyzed. Samples with low γH2AX (<25% of geminin-positive cells with γH2AX foci) or with <40 geminin-positive cells were not included in the analyses, due to insufficient endogenous DNA damage or tumor cells in the S/G2-phase of the cell cycle, respectively. The mouse anti-geminin antibodies used in this study are human-specific so that they did not cross-react with mouse stroma.

Patient-derived tumor cells were isolated from 24 PDX through combination of mechanic disruption and enzymatic disaggregation following a previously described protocol (16). Briefly, PDX tumors not bigger that 500 mm³ were freshly collected in DMEM/F12/HEPES (GIBCO) after surgery resection, minced using sterile scalpels and dissociated for a maximum of 90 minutes in DMEM/F12/HEPES (GIBCO), 1 mg/mL collagenase (Roche), 100 μg/mL hyaluronidase (Sigma-Aldrich), 5% BSA (Sigma-Aldrich), 5 µg/mL insulin, and 50 µg/mL gentamycin (GIBCO). This was followed by further dissociation using 0.05% trypsin (GIBCO),1 µg/mL dispase (StemCell technologies) and 1 mg/mL DNase (Sigma-Aldrich). Red blood cell lysis was done by washing the cell pellet with 1× red blood cell (RBC) lysis buffer containing ammonium chloride (Invitrogen). Then, cells were resuspended RPMI1640 with GlutaMAX medium (Gibco) supplemented with 2% of heat inactivated FBS (Gibco), 1 µL/mL ROCK inhibitor, 20 µL/mL B27, 0.02 µL/mL EFG, 0.2 µL/mL FGF10, 0.2 µL/mL FGF2. To obtain FFPE blocks from PDC cultures, cells were seeded at 2 × 10⁵ cells/mL in 6-well plates (BD Biosciences). After 24 hours, PDC were treated with vehicle (DMSO) or 2.5 µmol/L olaparib and incubated at 37°C in 5% of CO₂. Vehicle and olaparib-treated cells were harvested, fixed in 4% paraformaldehyde (PFA), followed by 70% ethanol, and embedded in paraffin.

For IC₅₀ analysis, cells (2 × 10³ cells/well) were seeded into 96-well plates (BD Biosciences) and were treated the following day with different concentrations of olaparib for 14 days. The treatments and media were refreshed every 2 days. Olaparib-sensitive and -resistant PDX were included in every batch of analysis. Cell viability was quantified with CellTiter Glo Luminiscent Cell Viability Assay (Promega) in an Infinite M200 PRO plate reader (TEKAN). Values at day 14 were subtracted with the mean background signal (no cells), normalized with values at day 0, relativized to controls (not treated cells) and plotted as the percentage inhibition against the log concentration of olaparib. IC₅₀ values (half maximal inhibitory concentration) were calculated performing a nonlinear regression. Each experiment was repeated twice with two technical replicates.

Data were analyzed with GraphPad Prism version 7.0 software (GraphPad). Bars represent the median of at least two technical replicates, unless otherwise stated. Shapiro–Wilk test was used to assess normality of data distributions. If the null hypothesis of normal distribution was not rejected, statistical tests were performed using paired and unpaired two-tailed t test (for two groups comparison of PDC and/or patient samples, and HRD score by Myriad in PDX). Otherwise, the nonparametric Mann–Whitney or Wilcoxon signed-rank test was used (for two groups comparison of RAD51 score and HRDetect score in PDX). The ROC curve and the AUC were calculated to estimate the prediction capacity of Myriad's myChoice HRD test and RAD51 scores to PARPi response. For AUC comparison, a two-sided bootstrap test was used by means of statistical package pROC in R software. In addition, sensitivity, specificity, positive (PPV), and negative (NPV) predictive values were calculated using the pre-established cut off. Cohen κ coefficient was calculated to estimate the agreement of RAD51 status between PDX and patient samples. To calculate the association between RAD51 score in PDX and in PDC, and between RAD51 in PDX and patient samples, a linear regression model was fitted to estimate the R² with 95% CI.

The main goal of this work is to analyze in a cohort of PDX models whether the accuracy to predict response to PARPi is different using mutations in HRR genes, genomic HRD tests, or the RAD51 assay. Our PDX panel (see below) is enriched with models having HRD, as >50% of models derived from patients with a germline HRR gene alteration. In addition, it is expected to find >70% of HRD-positive tumors by genomic scars since the vast majority of the PDX tumors are TNBCs according to immunohistochemistry characterization of estrogen and progesterone receptor and the HER2 status (26). The accuracy to predict PARPi response using HRR-gene mutations is expected to be 70% whereas the accuracy of RAD51 assay is estimated to be 92% (27, 28). To detect such a difference using a 1/2.5 ratio with at least 90% power and two-sided type I error of 0.05, at least 41 PDX samples with HRR-gene mutations and 103 PDX samples with RAD51 assay information are required. Sample size calculation is based on exact binominal distribution. Assuming that not all HRR-gene tests and RAD51 assays will provide informative results, an attrition rate of 5% is estimated, and consequently, a total of 109 PDX samples are needed.

Readers can find the genomics data acquired and used in this study in Supplementary Table S3.

One hundred and nine patient-derived models from three BRCA-associated cancers were established, namely from breast, ovarian, and pancreatic cancers, both from primary and metastatic tumors (Fig. 1; Supplementary Table S1; refs. 14, 15). The antitumor activity of the PARPi olaparib or niraparib was assessed in 96 PDX models derived from 65 TNBC, 24 hormone receptor-positive (HR+) breast cancer, 1 HER2-positive (HER2+) breast cancer, 4 HGSOC, and 2 patients with PaC (Supplementary Table S1). Upon PARPi treatment in vivo, the best response according to modified RECIST criteria was: complete response (CR) in 12, partial response (PR) in 10, and stable disease (SD) in 7 PDX models, totaling 29 of 109 sensitive models. Sixty-seven PDX models were PARPi-resistant [progressive disease (PD), as best response; Fig. 2A]. Additional 13 resistant models were generated from nine PARPi-sensitive PDX after prolonged exposure and steep progression to olaparib. Time to progression was not observed to be associated with the type of HRR alteration (Supplementary Fig. S1A). In total, 80/109 models were resistant to PARPi.

Forty of 96 (42%) PDX models did not present alterations in HRR genes and did not respond to PARPi treatment (Fig. 2A and B). The remaining 56 models harbored alterations in HRR genes and presented differential responses to PARPi (Fig. 2A and B). Pathogenic biallelic variants in BRCA1 were identified in 25 models (9 sensitive and 16 resistant) and BRCA2 pathogenic biallelic variants were detected in 13 models (10 sensitive and 3 resistant). BRCA1 epigenetic silencing, evaluated by lack of BRCA1 transcript by RNA-seq (cut off ≤1 tmp) and lack of functional BRCA1 nuclear foci by immunofluorescence (IF; cut off ≤10%) was found in 13 PDX (6 sensitive and 7 resistant). We identified biallelic frameshift mutations in PALB2 in three PARPi-sensitive models (PDX093, T298, ST897) and homozygous loss of the gene encoding the RAD51 paralog XRCC3 in a PARPi-sensitive PDX (HBCx14, labeled as “other HRR genes” in Fig. 2A and B; Supplementary Table S1). Taken together, our analyses identified that all the PARPi-responsive PDX models in the cohort had at least one alteration in the HRR genes BRCA1, BRCA2, PALB2, or XRCC3.

WES also revealed mutations that could explain mechanisms of resistance to PARPi in HRR-altered PDX (Fig. 2C). Three models (PDX093OR6, PDX332, and PDX405OR100) harbored mutations in PALB2 or BRCA2 that were compatible with HRR gene reversions. Three additional BRCA1-mutant models presented alterations in the 53BP1–shieldin pathway, which have been shown to confer PARPi resistance (HBCx28, PDX127, PDX230OR1; Supplementary Table S1; refs. 29–31). Interestingly, BRCA1 nuclear foci (cut off >10%) were detected in 7 of 25 BRCA1-mutant models, underscoring the high prevalence of re-expression of BRCA1 proteins lacking relevant domains in this panel, which is considered hypomorphic (Supplementary Table S1). The presence of BRCA1 proteins lacking relevant domains in tumors with mutations in the RING, exon 11, or the BRCT domains of BRCA1 has been associated with resistance to PARPi (Fig. 2C; Supplementary Table S1; refs. 32–34). In addition, BRCA1 nuclear foci were detected in 4 of 11 BRCA1-low models upon acquisition of PARPi-resistance. Finally, three models (PDX124OR, PDX377OR1, and PDX377OR2) were PARPi-resistant without substantial of HRR restoration, according to their levels of BRCA1 and RAD51 nuclear foci when compared with their sensitive counterparts (named as “RAD51-independent” in Fig. 2C). Altogether, our genomics, transcriptomics and BRCA1 protein foci analyses identified likely mechanisms of PARPi resistance in half of the HRR-altered, nonresponsive models.

We calculated the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for PARPi response of HRR gene mutations or low BRCA1 expression (mRNA or nuclear foci, Table 1). These two HRD biomarkers showed similar accuracy (67% and 69%, respectively). In summary, our data confirmed that genetic and epigenetic alterations in HRR genes had limited accuracy predicting PARPi sensitivity.

We evaluated the association between the genomic scar Myriad myChoice HRD and PARPi response. Genomic HRD was quantified in a subset of 41 PDX models, 11 of which harbored a mutation in BRCA1/2 or PALB2. All the HRR gene–mutated samples were HRD-positive (cut off ≥42; Supplementary Fig. S1B; Supplementary Table S1). In addition, the test identified as HRD-positive nine tumors lacking BRCA1 expression (BRCA1 low), of which, two were known to previously harbor BRCA1 promoter hypermethylation (named as “WT with history of HRD” in Supplementary Fig. S1B; ref. 35). All PARPi-sensitive PDX were HRD positive and the median of HRD values was higher in PARPi sensitive than in PARPi resistant models (Fig. 3A, median of 67 vs. 36, P = 0.0002). Nonetheless, 12 of 29 PARPi-resistant PDX were also HRD positive. Accordingly, the genomic HRD test showed a limited accuracy for predicting PARPi responses (71%), similar to HRR-gene mutations or assessment of BRCA1 mRNA expression or BRCA1 nuclear foci (Table 1).

We also quantified the WGS-derived HRD signature named HRDetect in an independent subset of 15 PDX, 10 of which were derived from BRCA1/2 mutation carriers and 11 from metastatic biopsies (Supplementary Table S1). All PDX derived from patients with BRCA1/2-associated breast cancers had high HRDetect values (cut off >0.7; ref. 6). The GIS score in these models was concordant with HRDetect (Supplementary Fig. S1C). As observed for the Myriad myChoice HRD scores, all PARPi-sensitive tumors had high HRDetect values, but 8 of 11 PARPi-resistant tumors also had high HRDetect values (Supplementary Fig. S1D). In summary, these data show that genomic HRD tests provide very high sensitivity and NPV but limited specificity and PPV to predict PARPi response.

We further aimed to investigate the functional HRD status of this PDX cohort, scoring the percentage of tumor cells with RAD51 nuclear foci in the S–G₂ phase of the cell cycle (geminin-positive; refs. 14, 15). When comparing RAD51 with HRR-gene alterations, we noted that HRR-altered models had significantly lower levels of RAD51 values than HRR-WT PDX (median of 16% vs. 50%, P < 0.0001; Supplementary Fig. S1E), albeit 61% of HRR-gene altered tumors had high RAD51 scores. Similarly, 54% of tumors identified as positive for genomic HRD presented high RAD51 scores (Fig. 3B).

Next, we explored the potential of the RAD51 score to predict PARPi response. PARPi-sensitive models showed a significantly lower percentage of RAD51-positive cells than the PARPi-resistant ones (median of 1% vs. 38% respectively, P < 0.0001; Fig. 3C). Moreover, we observed an increase of the RAD51 score in the 13 models of acquired PARPi resistance derived in laboratory conditions, compared with their sensitive counterparts (median of 33% vs. 1%, P = 0.0006; Supplementary Fig. S1F). This result highlights the dynamic nature of the RAD51 test.

Using the pre-defined cut off for RAD51 (RAD51 ≤10%; ref. 9), the RAD51 assay demonstrated high sensitivity (90%) and specificity (98%), along with high PPV (93%) and NPV (96%; Table 1). In more detail, the RAD51 test predicted PARPi response in PDX harboring BRCA1, BRCA2, and PALB2 mutations (accuracy 92%; Supplementary Fig. S1G) and in PDX without mutation in these genes (accuracy 98%; Supplementary Fig. S1H). The accuracy of RAD51 to predict PARPi response was significantly higher than the obtained with the genomic HRD test (95% vs. 71%, P < 0.001). Of note, the discordant models according to the RAD51 test exhibited an intermediate response to PARPi (Supplementary Figs. S2A–S2C). A comparative ROC analysis was conducted for RAD51 scores and Myriad myChoice HRD test in the subset of 41 samples and revealed that the ROC AUC for the RAD51 assay was superior than for the genomic HRD score (0.97 vs. 0.81, P = 0.03; Fig. 3D). Altogether, these data validate the pre-defined RAD51 assay score cut off and demonstrate that assessment of RAD51 nuclear foci by immunofluorescence is the most accurate and dynamic HRD biomarker for predicting PARPi response (Table 1).

We established cisplatin response in a subset of 56 PDX, of which, 51 were derived from breast cancers (Fig. 4A). As expected, most of the PARPi-sensitive models were also sensitive to platinum (18/21, 86%) and none of the PARPi-resistant PDX restoring BRCA1 function were platinum sensitive (0/8, 0%). Instead, two 53BP1-dependent, PARPi-resistant PDX were platinum sensitive (2/2, 100%; ref. 36).

We observed that platinum response provided lower accuracy to predict PARPi response than the RAD51 score (54% vs. 95%, Table 1). Regarding RAD51 scores and cisplatin response, we observed that cisplatin-sensitive PDX models had overall lower values of RAD51 than cisplatin-resistant PDX (median of 12% vs. 21%, P = 0.02, Fig. 4B). However, 16 of 31 of the cisplatin-sensitive models presented high RAD51 scores, which suggests that these models carry deficiencies in other DNA repair pathways involved in repair of platinum-induced damage (e.g., nucleotide excision repair or Fanconi anemia pathway). Accuracy of the RAD51 assay to predict cisplatin response was similar to HRR-gene mutations or to BRCA1 epigenetic silencing (64%, 61%, and 63%, respectively; Table 2). These results suggest that the RAD51 assay could help to identify patients who are likely to respond to platinum-based treatments with a similar accuracy of other HRD tests currently used in clinical practice.

The HRD status by RAD51 in PDX was highly concordant with the corresponding patient's FFPE tumor sample that was taken at the same time-point when the PDX model was established (Cohen κ = 0.93, Fig. 5A). In agreement, the PDX response to PARPi or platinum was also highly concordant to the patient's response to the respective drug (30 of 36, 83% concordance; Supplementary Table S2).

Patient-derived tumor organoids from HGSOC are being established to assess drug responses to targeted drugs in a real-time manner (37). Therefore, we further assessed the ability of patient-derived tumor cells (PDC) to recapitulate the HRD status of the PDX and the in vivo responses to PARPi. In a cohort of 14 models, the RAD51 score in PDC was highly concordant with the score in PDX, both when the models were PARPi-treated (κ coefficient = 0.68) or left untreated (κ coefficient = 0.74, Fig. 5B; Supplementary Fig. S3). Concordantly, treatment with olaparib resulted in lower IC₅₀ values in 24 PDC derived from PARPi-sensitive PDX than from resistant ones (10² nmol/L vs. 10^2.9 nmol/L, P = 0.017, Fig. 5C and D). In summary, we demonstrated that both PDX and PDC recapitulate patient's HRD status and in vivo response to PARPi.

In the era of precision oncology, there is an unmet medical need for optimized use of PARPi and platinum-based chemotherapy. The use of validated HRD tests in clinical practice may fill this gap. Ideally, this test should offer a fast turnaround time, be accurate and of low cost, and be close to the point of care. In this work, we validated the predictive value of the RAD51-immunofluorescence test for PARPi response in a panel of patient-derived tumor models that recapitulated patients’ characteristics and response to PARPi or platinum. Compared with genetic and genomic HRD tests, the RAD51 test exhibited greater PPV (93% vs. 43–50%), implying a better capacity to identify patients with PARPi-sensitive tumors and differentiate from those that become PARPi-resistant after restoration of functional HRR. Alike genomic HRD tests, RAD51 identified HRD in tumors with epigenetic silencing of BRCA1, as evidenced by concomitant low expression of BRCA1 mRNA expression or low BRCA1 nuclear foci. Importantly, the RAD51 test captured the dynamic nature of HRD as shown in paired models with acquired resistance to PARPi. Regarding platinum response, the RAD51 test showed similar accuracy to HRR-gene mutations (64% vs. 61%).

In our PDX cohort, 58% of models harbored an HRR alteration, mainly driven by genetic alterations in BRCA1/2 or epigenetic silencing of BRCA1. The preclinical response rate to PARPi among these models was only of 42%, highlighting the limited PPV of these selection biomarkers. We identified the potential mechanism of resistance in half of the resistant models, encompassing the restoration of HRR-gene function in 24%, restoration of end resection mediated by the 53BP1 pathway in 4% or RAD51-independent mechanisms in 4%. RAD51-independent mechanisms of PARPi sensitivity or resistance represent a limitation for the accuracy of the RAD51 test. As an example, enhanced PARP trapping causes PARPi toxicity irrespective of HRD, such as in tumors with impaired ribonucleotide or base excision repair due to defects in RNaseH2 or XRCC1, respectively (38, 39). Impairment of enzymes involved in other base damage repair pathways, for example, loss of ALC1 nucleosome remodeling, also confers PARPi sensitivity (40, 41). The clinical relevance of these mechanisms is yet to be established. For 26% of the PDX in our cohort, the mechanism of resistance was not revealed by WES, but consistent with restoration of HRR by RAD51 nuclear foci formation.

The lack of PPV of the tests based on detection of HRR alterations (gene mutations or epigenetic silencing) lies upon the heterogeneity of the HRD nature. Also, even though genomic HRD tests encompass the different sources of HRD, they cannot distinguish HRR restoration. In contrast, a dynamic functional biomarker alike RAD51 can capture this heterogeneity and dynamism. The discordance between genomic HRD tests and RAD51 foci is observed in half of the PDX, especially in models harboring BRCA1 epigenetic silencing, likely reflecting the “soft” status of the phenotype. This observation is in line with the TBCRC009 and TNT trials for platinum (42, 43). In summary, a functional test like RAD51 foci shows an improved performance because it captures the different sources of HRD as well as its dynamic evolution.

In patients with metastatic breast cancer, the PARPi olaparib and talazoparib have been approved for patients with a germline BRCA1 (gBRCA1) or germline BRCA2 (gBRCA2) pathogenic variant. Nevertheless, the overall response rate is within the range of 60% (44, 45). In the context of a known germline condition, a functional test that measures the tumor status of HRD before treatment initiation would likely improve patient selection for PARPi monotherapy. This would allow to optimize the prescription of effective drugs in patients who have restored HRR in their tumors. A similar scenario is expected for patients with germline pathogenic genetic alterations in PALB2 who are also sensitive to PARPi (27, 46). In addition, efforts are also being invested in expanding the patient population who may benefit from PARPi beyond those harboring germline HRR-gene pathogenic variants with metastatic cancer. Given the positive safety profile of PARPi and preliminary efficacy data in the early disease setting (28, 47), it is expected that PARPi will be incorporated as part of the neoadjuvant and adjuvant treatment regimens (48). In this regard, data supporting the use of genomic HRD tests or RAD51 as predictive biomarkers of PARPi response in breast cancer is encouraging, and in early breast cancer, and analysis in larger cohorts is awaited (20, 49).

Regarding platinum response and breast cancer, genomic scars have failed to identify patients who may benefit both in the early and metastatic settings, and gBRCA1/2-mutations predicted platinum response exclusively in the metastatic setting (43, 50). Interestingly, despite cisplatin sensitivity may occur beyond HRD (51), our preclinical data showed that the RAD51 assay may help select patients who benefit from platinum treatment, thus giving a “target” therapy option to gBRCA1/2 WT patients. In untreated TNBC, RAD51 independently predicted clinical benefit from adding carboplatin to neoadjuvant chemotherapy (22). Further studies are needed to confirm the clinical validity of the RAD51 assay at predicting long-term platinum benefit, and to verify if the test could predict response to treatments with similar mechanism of action.

Functional biomarkers of HRD are also needed for the management of HGSOC, advanced prostate cancer, or PaC. Despite quality of tumor genetic testing and interpretation of variants’ effect have greatly improved in the recent years, there is still a scarcity of implementation of tumor genetic testing beyond the ovarian cancer population. Regarding platinum-sensitive HGSOC, PARPi have been approved as first- and second-line maintenance therapy (1, 52–54). Recently, the European Medicines Agency (EMA) approved olaparib in combination with bevacizumab (an anti-VEGF monoclonal antibody) as first-line therapy in patients with HRD-positive platinum-sensitive HGSOC, using an HRD diagnostic test with demonstrated clinical validity (1, 8, 55–57). This setting provides an opportunity to clinically validate the RAD51 assay. Regarding prostate cancer, the PARPi olaparib and rucaparib are now approved for men with metastatic disease (1, 58), but significant intrapatient variability in clinical outcomes has been observed, particularly for those patients with alterations in genes beyond BRCA2. A functional assay such as RAD51 can be relevant to assess the clinical value of less-frequent alterations in other DNA damage repair genes, for example, PALB2, and guide patient stratification (21). Finally, olaparib is the first targeted therapy that has been approved for gBRCA1/2-mutated pancreatic tumors that previously responded to platinum salts (59). In this context, the RAD51 assay could be helpful to: (i) avoid platinum-based chemotherapy as selection biomarker, thus reducing toxicities for patients who may not benefit from them; (ii) select patients that needed to otherwise be tested for gBRCA1/2 status, considering the low incidence of the mutations in this subset of patients; and (iii) increase the number of patients who may benefit from PARPi beyond those with gBRCA1/2 mutations.

We foresee that the RAD51 test will have some limitations in relationship with PARPi response: (i) it will not identify tumors whose sensitivity to PARPi lies on a DNA repair deficiency outside of the HRR core pathway (i.e., due to alterations in ATM, RNaseH2, XRCC1, ALC1); and (ii) it will not identify secondary PARPi-resistance when the mechanism does not involve restoration of HRR (i.e., replication fork stabilization). In summary, the RAD51 assay could help to identify (i) patients harboring gBRCA1/2 or gPALB2 mutations who are likely to be PARPi-resistant and (ii) patients without a germline condition in BRCA1/2 or PALB2 with HRD-positive tumors who are likely to be PARPi-sensitive. Using PDX we confirmed the superior predictive value of an immunofluorescence-based test detecting RAD51 nuclear foci in preclinical models treated with PARPi when compared with HRR-gene mutations, genomic HRD tests and platinum sensitivity. Given the supportive clinical data (20, 22, 49), we propose to continue the analytical validation and the clinical qualification of the RAD51 assay in larger cohorts of patients with breast cancer treated with PARPi or platinum salts.

B. Pellegrino reports grants from Roche during the conduct of the study, other support from Pfizer and Novartis, and personal fees from Merk outside the submitted work. A. Llop-Guevara reports a patent for PCT/EP2018/086759 pending. F. Pedretti reports grants from “La Caixa” Foundation (ID 100010434) during the conduct of the study. G. Villacampa reports personal fees from MSD, AstraZeneca, and Pierre Fabre outside the submitted work. A. Degasperi reports a patent for HRDetect pending to CRUK. J.V. Forment reports that he is a full-time employee and shareholder of AstraZeneca. M.J. O'Connor reports other support from AstraZeneca outside the submitted work; also has a patent for RAD51 foci assay as a patient selection approach pending. Y. Zhou reports other support from GSK (formerly TESARO) outside the submitted work. A. Musolino reports grants from Roche, grants and personal fees from Lilly, and personal fees from Eisai, Seagen, and Daiichi-Sankyo outside the submitted work. C. Caldas reports grants from AstraZeneca, Genentech, Roche, Servier, and Varsity, and personal fees from Illumina and AstraZeneca during the conduct of the study. S. Nik-Zainal reports a patent for HRDetect pending to CRUK. P.G. Nuciforo reports personal fees from Novartis, Bayer, and MSD Oncology outside the submitted work. V. Peg reports personal fees from Roche, AstraZeneca, Sysmex, MSD, Exact Sciences, and Bayer outside the submitted work. M. Espinosa-Bravo reports personal fees from Roche Ldt and Sysmex outside the submitted work. T. Macarulla reports personal fees from SOBI - Swedish Orpahn Biovitrum AB, Ability Pharmaceuticals SL, and Aptitude Health; personal fees and other support from AstraZeneca, Incyte, and Servier; personal fees from Basilea Pharma, Baxter, BioLineRX Ltd., Celgene, Eisai, Ellipses, Genzyme, Got It Consulting SL, Hirslanden/GITZ, Imedex, Ipsen Bioscience, Inc., Janssen, Lilly, Marketing Farmacéutico & Investigación Clínica, S.L., MDS, Medscape, Novocure, Paraxel, PPD Development, Polaris, QED Therapeutics, Roche Farma, Sanofi-Aventis, Scilink Comunicación Científica SC, Surface Oncology, and Zymeworks outside the submitted work. A. Oaknin reports personal fees from Agenus, Clovis Oncology, Inc., Corcept Therapeutics, Deciphera Pharmaceutical, Eisai Europe Limited, EMD Serono, Inc., F. Hoffmann-La Roche, GlaxoSmithKline, Immunogen, KL Logistics, Medison Pharma, Merck Sharp & Dohme de España, Mersana Therapeutics, Novocure GmbH, Pharma Mar, prIME Oncology, ROCHE FARMA, Sattucklabs, Sutro Biopharma, Inc., Immunogen, KL Logistics (AMGEN), Medison Pharma, Merck Sharp & Dohme, Mersana Therapeutics, Novocure, and prIME Oncology, Sattucklabs, Sutro Biopharma, Inc., Genmab; personal fees and other support from AstraZeceneca and Pharma Mar, ROCHE FARMA; grants from Abbvie Deutschland Gmbh & Co Hg, Ability Pharmaceuticals, Advaxis, Agenus, Aprea Therapeutics AB, AstraZeneca AB, Beigene USA, Inc., Belgian Gynaecological Oncology Group (BGOG), Bristol-Myers Squibb International Corporation (BMS), Clovis Oncology, Corcept Therapeutics, Eisai, F. Hoffmann-La Roche, Immunogen, Iovance Biotherapeutics, Lilly, Medimmune, Merck Healthcare, Merck Sharp & Dohme, Millennium Pharmaceuticals, Mundipharma Research, Novartis Farmacéutica, Regeneron Pharmaceuticals, Seagen, Seattle Genetics, Sutro Biopharma, Tesaro, and Werastem outside the submitted work. J. Mateo reports grants, personal fees, and nonfinancial support from AstraZeneca and Pfizer Oncology; personal fees from Janssen, Astellas, and Guardant Health; and personal fees and nonfinancial support from Roche outside the submitted work. J. Arribas reports grants from Roche, Synthon/Byondis, and Molecular Partners; and grants and personal fees from Menarini during the conduct of the study; also has a patent for P200801652 licensed and with royalties paid and a patent for EP20382457.8 licensed and with royalties paid. R. Dienstmann reports personal fees from Roche, Amgen, Libbs, Bayer, MSD, Servier, Sanofi, and Boehringer-Ingelheim; grants from Merck, BMS, and Pierre-Fabre outside the submitted work. M. Bellet reports other support from Pfizer, Novartis, and Lilly outside the submitted work. M. Oliveira reports grants, personal fees, and nonfinancial support from Roche and Novartis; grants and personal fees from AstraZeneca, SeaGen, GlaxoSmithKline, PUMA Biotechnology; personal fees from Gilead, iTEOS, and Guardant Health; grants from Genentech and Immunomedics, Sanofi, Zenith Epigenetics, and Boehringer Ingelheim; personal fees and nonfinancial support from Pierre Fabre and Eisai outside the submitted work. C. Saura reports personal fees from AstraZeneca, AX'Consulting, Byondis B.V., Daiichi Sankyo, Eisai, Exact Sciences, Exeter Pharma, F. Hoffmann - La Roche Ltd., ISSECAM, Medical Statistics Consulting, MediTech, Merck Sharp & Dohme, Novartis, Pfizer, Philips, Piere Fabre, PintPharma, Puma, Sanofi, SeaGen, and Zymeworks; personal fees and other support from Roche Farma; grants from Aragon Pharmaceuticals, AstraZeneca AB, AstraZeneca Farmacéutica Spain, Bayer Pharma, Boehringer Ingelheim España, Bristol-Myers Squibb (BMS), Cytomx Therapeutics, Daiichi Sankyo, F. Hoffmann-La Roche Ltd., Genentech, German Breast Group Forchungs, Glaxosmithkline, Innoup Farma, International Breast Cancer Study Group (IBCSG), Lilly, Macrogenics, Medica Scientia Innovation Research, Menarini Ricerche, Merus, Millennium Pharmaceuticals, Novartis Farmacéutica, Pfizer, Puma Biotechnology, Roche Farma, Sanofi-Aventis, Seattle Genetics, and The Netherlands Cancer Institute-Antoni van Leeuwenhoek Ziekenhuis outside the submitted work. J. Balmaña reports grants from AstraZeneca, personal fees and nonfinancial support from AstraZeneca and Pfizer during the conduct of the study; also has a patent for EP17382884.9 issued. V. Serra reports grants and personal fees from AstraZeneca; grants from Tesaro; and personal fees from Abbvie during the conduct of the study; also has a patent for WO2019122411A1 issued. No disclosures were reported by the other authors.

B. Pellegrino: Formal analysis, funding acquisition, methodology, writing–original draft, writing–review and editing. A. Herencia-Ropero: Formal analysis methodology, writing–review and editing. A. Llop-Guevara: Conceptualization, data curation, supervision, funding acquisition, methodology, writing–review and editing. F. Pedretti: Formal analysis, writing–review and editing. A. Moles-Fernández: Formal analysis. C. Viaplana: Data curation. G. Villacampa: Formal analysis, methodology, writing–review and editing. M. Guzman: Methodology. O. Rodriguez: Methodology. J. Grueso: Methodology. J. Jiménez: Resources. E.J. Arenas: Resources. A. Degasperi: Formal analysis, writing–review and editing. J.M. Dias: Formal analysis, writing–review and editing. J.V. Forment: Data curation. M.J. O'Connor: Resources. O. Déas: Formal analysis. S. Cairo: Data curation. Y. Zhou: Resources, data curation. A. Musolino: Funding acquisition, writing–review and editing. C. Caldas: Resources, funding acquisition, writing–review and editing. S. Nik-Zainal: Resources, funding acquisition, writing–review and editing. R.B. Clarke: Resources, funding acquisition, writing–review and editing. P.G. Nuciforo: Resources. O. Díez: Data curation. X. Serres-Créixams: Resources, data curation, software, writing–review and editing. V. Peg: Resources. M. Espinosa-Bravo: Resources. T. Macarulla: Resources. A. Oaknin: Resources. J. Mateo: Writing–review and editing. J. Arribas: Resources, supervision, funding acquisition, writing–review and editing. R. Dienstmann: Resources, writing–review and editing. M. Bellet: Resources. M. Oliveira: Resources. C. Saura: Resources. S. Gutiérrez-Enríquez: Data curation, formal analysis, funding acquisition, writing–review and editing. J. Balmaña: Conceptualization, resources, supervision, funding acquisition, writing–review and editing. V. Serra: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, project administration, writing–review and editing.

The authors thank the patients for providing study samples. The authors acknowledge the CELLEX foundation for providing research equipment and facilities. They thank CERCA Program/Generalitat de Catalunya for institutional support. The authos acknowledge funding from AstraZeneca and Tesaro supporting this study (to V. Serra). The GHD-Pink program, the FERO Foundation, and the Orozco Family provided funding for establishment of the PDX models at VHIO. This study has also been supported by the Catalan Agency AGAUR-FEDER (2017 SGR 540, 2019PROD00045; to V. Serra and A. Llop-Guevara). V. Serra, J. Balmaña, and S. Gutiérrez-Enríquez received funds from the Instituto de Salud Carlos III (ISCIII), an initiative of the Spanish Ministry of Economy and Innovation, partially supported by European Regional Development FEDER Funds: grants PI12/02606, PI17/01080, and PI19/01303; PI20/00892, CP14/00228, and CPII19/00033. The study is also the result from a Transcan-2 effort (AC15/00063), the Spanish Association of Cancer Research (AECC, LABAE16020PORTT), and from the SPORE in Breast Cancer (NIH P50CA098131). La Caixa Foundation and the European Institute of Innovation and Technology/Horizon 2020 (CaixaImpulse; LCF/TR/CC19/52470003) provided funds to A. Llop-Guevara. B. Pellegrino was supported by ESMO with a Clinical Translational Fellowship aid supported by Roche. Any views, opinions, findings, conclusions, or recommendations expressed in this material are those solely of the authors and do not necessarily reflect those of ESMO or Roche. A. Herencia-Ropero received funding from Generalitat de Catalunya (PERIS SLT017/20/000081). A. Llop-Guevara received funding from Generalitat de Catalunya (PERIS SLT002/16/00477) and AECC (INVES20095LLOP). F. Pedretti received a fellowship from “La Caixa” Foundation (ID 100010434; LCF/BQ/DI19/11730051). This work was also supported by Breast Cancer Research Foundation (BCRF-19-08), Instituto de Salud Carlos III project AC15/00062, and the EC under the framework of the ERA-NET TRANSCAN-2 initiative co-financed by FEDER, Instituto de Salud Carlos III (CB16/12/00449 and PI19/01181), and AECC (to J. Arribas). M. Espinosa-Bravo reports grants from FERO Foundation and from ISCIII. J. Mateo received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement 837900. S. Nik-Zainal, J. Dias, and A. Degasperi are supported by Cancer Research UK (CRUK) Advanced Clinician Scientist Award grant C60100/A23916, CRUK Pioneer Award, CRUK Grand Challenge Award grant C60100/A25274, and the National Institute of Health Research (NIHR) Cambridge Biomedical Research Centre grant BRC-125-20014. The views expressed are those of the authors and not necessarily those of the NIHR or the Department of Health and Social Care.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.