TNBC and Ovarian cancer PDX models establishment was as described (36, 37). PDX models used are described in Table 1 . The study was reviewed and approved by the ethics committees for animal experimentations of the University of Montpellier (CEEA-LR-12028). PDX models were established from fresh tumor fragments obtained from the Pathology Department at the Comprehensive Cancer Center of Montpellier (ICM) after informed consent of the patients. Establishment of PDX models was reviewed and approved by the institutional review board. Approximately 50 mm³ PDX fragments were grafted subcutaneously into the flank of 3-4week old Swiss-nude female mice (Charles Rivers, Saint-Germain-sur-l’Arbresle, France). The present study comprised three experimental arms; vehicle, olaparib and Carboplatin (CBP) treatment comprising 6 to 8 mice per arm. When median tumor volume reached 100-150mm³, mice were randomly distributed in the two arms and treatment was started. Olaparib (Lynparza, AstraZeneca) was administered orally 5 times/week for 5 weeks at 100 mg/kg. CBP (Accord Healthcare, Middlesex, UK) was administered by intra-peritoneal (IP) injection twice per week for 4 weeks at 50 mg/kg CBP. At treatment end, mice were euthanized to collect tumor samples for further biochemical (RNA and proteins) or histological analyses. Some mice were kept for tumor volume monitoring.

CpG methylation patterns at the BRCA1 promoter were determined using the MS-PCR assay as previously described (28).

For each PDX sample, the genomic profile was established by using aCGH onto high-resolution 4 × 180 K CGH-microarrays (SurePrint G3-Human CGH-Microarray, Agilent Technologies). Human female DNA was used as reference (G152A, Promega, Madison, WI, USA). Both experimental and analytic methods have been previously described (38). All probes for aCGH were mapped according to the Genome Reference Consortium Human Build 37 (CGCh37/Hg19; https://www.ncbi.nlm.nih.gov/assembly/GCF_000001405.13/). We used two different threshold values (log2 ratio > 0.5 and 1.0) to distinguish low- (gain/) from high-(amplification) and (log2 ratio < -0.3 and -1) to distinguish simple loss from deletion CAN (Copy Number Alterations), respectively. Percentage of altered genome was the number of probes above the threshold divided by the total number of probes for autosomal chromosomes.

For each tumor, to evaluate genetic instability, we quantified the activity of the 17 copy number signatures described (20) with the R package CINSignatureQuantification.

For each tumor, a HRD score (HRDaCGH score), based on losses of heterozygosity (LOH), was calculated as previously described (39). A score ≥ 10 was considered as HRD-high. The percentage of genome altered was calculated as the sum of altered probes divided by the total number of probes after removing sexual chromosomes.

Mutation profiles were established by using targeted-NGS panel of 559 genes commonly mutated in breast cancer (40). The list of genes targeted is available in the Supplementary Information. The sequence data were aligned to the human reference genome (UCSC hg19) using Burrows–Wheeler Aligner (41). Tumor samples were sequenced at an average depth of 851× (range, 520 to 1055) for the targeted regions. Bam files were processed as described (40). Single nucleotide variants (SNVs) calling was performed with a consensus approach using 8 variants callers for the SNV (Freebayes) (42), HaplotypeCaller (43), LoFreq (44), Mutect2 (45), Pisces (46), Platypus (47), VarDict (48) and Varscan2 (49) and 10 variants callers for the indel (FreeBayes (42), HaplotypeCaller (43), LoFreq (44), Mutect2 (45), pindel (50), Pisces (46), Platypus (47), Scalpel (51), VarDict (48) and Varscan2 (49)). Variants called by less than 5 variants callers were filtered out. Then, the filtered variants were annotated with the Annotate Variation Software (ANNOVAR, version 2013-11-12). Known variants found in dbsnp129 and dbsnp137 with a variant allele frequency (VAF) superior to 1% (1000 G or ESP6500) were removed. Finally, low frequency SNVs and indels that were suspected to be false positives were systematically inspected with IGV version 2.3.32 (52, 53). Mutations were classified as “neutral” or “damaging” using the majority rule of predictor software (provided by dbnsfp: Sift, Polyphen2, LRT, MutationTaster, MutationAssesor, FATHMM, RadialSVM, LR) as previously described (54). A “recurrent” mutation, also called “hot spot”, was defined as being found more than 10 times in the Catalogue of Somatic Mutations in Cancer (COSMIC V68) database (http://cancer.sanger.ac.uk/cosmic).

Total RNA was isolated from cell lines lysed in TRIzol (Invitrogen, Fisher Scientific, Illkirch-Graffenstaden, France), while PDX tumors were lysed using Lysing Matrix D (MP Biomedicals™, Doornveld, France). Subsequently, the RNA was extracted using the RNeasy Kit (Qiagen, Les Ulis, France) following manufacturer instructions. cDNAs were synthesized from 1μg of total RNAs using random hexamers and SuperScript III Reverse transcription (Invitrogen, Fisher Scientific, Illkirch-Graffenstaden, France). Real-time qPCR was performed on a LightCycler 480 SW 1.5 apparatus (Roche, Meylan, France) with ONEGreen^® FAST QPCR PREMIX (Ozyme, Saint Cyr l’Ecole, France) and designed human specific primers ( Supplementary Table 1 ). Results were quantified with a standard curve generated by serial dilutions of a reference cDNA preparation. GAPDH transcripts were used for normalization. The fold change in gene expression was calculated as: Fold change = 2^-ΔΔCT.

SUM159 and SUM149 TNBC cell lines a generous gift from Dr S Ethier (MUSC, Charleston, SC), were maintained in Ham’s F-12 medium (Gibco™, Fisher Scientific, Illkirch-Graffenstaden, France) supplemented with 5% FBS, 10 µg/ml insulin, 1µg/ml hydrocortisone and 1% Antibiotic-antimycotic (100X) (Gibco™, Fisher Scientific, Illkirch-Graffenstaden, France). UWB1.289PT cell line was obtained from the American Type Culture Collection (ATCC) and maintained in the 50% RPMI-1640 (Gibco™), 50% MEGM (MEGM Bullet Kit; CC-3150, Lonza, Basel, Switzerland) and supplemented with 3% FBS, 1% Antibiotic-antimycotic (100X) (Gibco™, Fisher Scientific, Illkirch-Graffenstaden, France). HCC38 cell line was obtained from ATCC and maintained in RPMI-1640 supplemented with 10% FBS and 1% Antibiotic-antimycotic.

For CRISPR-Cas9 generation of KO clones, SUM159PT cells were first transduced with a plasmid vector containing doxycycline inducible lentiviral expression of SpCas9. Lentiviral transduction was performed on 70% confluent cell cultures. Viral particles were added in the fresh medium containing 8µg/ml polybrene. After 16h the medium was changed and 2µg/ml puromycin added for cell selection for at least 5 days. Next, the cells were transduced as described above with two lentiviral plasmid vectors for the expression of sgBRCA1 (kind gift from Yea-Lih Lin, IGH, Montpellier). After lentiviral transduction, cells were selected with 400µg/ml G418 for 10 days and Cas9 expression was induced by treating a population of cells with 1µg/ml of doxycycline for 6 days. The cells were then cloned and clones verified for the KO by Sanger sequencing and western blot. All cell lines and selected clones were genetically typed by Eurofins Genomics cell line authentification (Eurofins Genomics, Les Ulis).

Protein extracts were prepared by lysing either tumor tissue or cell line pellets on ice for 30 min in Tris-HCl pH7.4 50mM, NaCl 100mM, NaF 50mM, β-glycerophosphate 40mM, EDTA 5mM, Triton X100 1%, Aprotinin 10mg/ml, PMSF 100mM, Leupeptin 1mM, Pepstatin 1mM, followed by a short centrifugation to pellet debris. Protein concentrations were measured using the BCA kit (Fisher Scientific, Illkirch-Graffenstaden, France) SDS-PAGE gel electrophoresis was done on 30µg protein samples subsequently transferred onto nitrocellulose membranes (Amersham, Velizy-Villacoublay, France) and incubated overnight at 4°C with the primary antibody. Antibodies used are listed in a separate section. Membranes were then washed and incubated with the appropriate secondary antibody in 5% non-fat dry milk in PBST for 2h at room temperature and revealed by incubation with Chemiluminescent HRP Substrate (Sigma Aldrich, Saint Quentin Fallavier, France).

For cell lines, cells were grown on 12mm diameter slides cover slips in 24 well-plate for 24h, then drugs were added at the predetermined IC50 concentration. After 24h drugs were washed off and cells prepared as described below. For tumor tissues, 6μm cryosections were prepared from OCT embedded deep frozen tissue and mounted on Fisherbrand™ Superfrost™ Plus Microscope Slides (Fisher Scientific, Illkirch-Graffenstaden, France) and stored at -80°C until used. Cells and tumor sections were sequentially subjected to mild extraction (0.4% Triton in PBS, 5min in cold), fixation (4% PFA diluted in PBS) and blockage/permeabilization (3% BSA + 0.2% Triton in PBS, 1 hour at room temperature), incubated overnight at 4°C with the primary antibody (diluted in 3% BSA + 0.2% Triton in PBS), then with the secondary antibody (diluted in 3% BSA + 0.2% Triton in PBS, 1h at room temperature). Between each step, slides were washed 3 times with PBS. Tumor cryosections were immersed 0.1% SBB (Sigma Aldrich, Saint Quentin Fallavier, France) and 70% ethanol for 20 minutes at room temperature to reduce tissue autofluorescence and subsequently washed three times for 5 minutes PBS with 0.02% Tween 20. Stained sections or cell were counterstained with DAPI (Fisher Scientific, Illkirch-Graffenstaden, France) to stain the nuclei and cover slips were mounted with MWL4-88 (Citifluor, CliniSciences, Nanterre, France) and stored at 4°C. Antibodies used are described in the Antibody section.

Immunofluorescence images were acquired using Zeiss microscope with a 63X-immersion oil lens and generated using Zeiss Blue software. RAD51, BRCA1, 53BP1 and γH2AX nuclear foci were scored using the CellProfiler image analysis software (version 2.2.0, Broad Institute). At least three biological replicates of each model (both vehicle- and olaparib-treated) were analyzed. Cells presenting >5 foci/nucleus for RAD51, BRCA1, 53BP1 or γH2AX were considered positive and tumors presenting >10% of positive cells were scored positive for the given marker.

Immunofluorescence; rabbit anti-RAD51 PC130 1:300 (Merck Millipore Sigma Aldrich, Saint Quentin Fallavier, France), rabbit anti-geminin 52508 1:200 (CST OZYME, Saint Cyr l’Ecole, France), mouse anti-BRCA1 sc-6954 1:100 (SCBT, Heidelberg, Germany), mouse anti-γ-H2AX (H2-3F4, kind gift from Dr. Mustapha Oulad-Abdelghani, MAB-IGBMC Illkirch-Graffenstaden, 1:4000), rabbit anti-53BP1 NB100-304 1:500 (Bio-techne LTD, Abington, UK). Secondary antibodies; goat anti-mouse Alexa Fluor 488 (Abcam ab150113, 1:1000), goat anti-rabbit Alexa Fluor 555 ab150078 1:1000 (Abcam, Cambridge, UK).

Western blotting; BRCA1 9010 1:500 (CST OZYME, Saint Cyr l’Ecole, France), BRCA2 A303-434A 1:1000 (Bethyl OZYME, Saint Cyr l’Ecole, France), PARP1 WH0000142M1 1:1000 (Sigma Aldrich, Saint Quentin Fallavier, France), RAD51 8875, 1:1000 (CST OZYME, Saint Cyr l’Ecole, France) and alfa tubulin T9026 1:20000 (Sigma Aldrich, Saint Quentin Fallavier, France); secondary antibodies goat anti-mouse-HRP 70745 1:10000 (CST OZYME, Saint Cyr l’Ecole, France) and goat anti-rabbit-HRP 7076 1:10000 (CST OZYME, Saint Cyr l’Ecole, France).

To test the relative sensitivity of TNBC with silenced BRCA1 due to promoter hypermethylation (designated hereafter BRCA1-Me) to the PARP inhibitor olaparib and assess the impact of neoadjuvant treatment (NACT) on olaparib sensitivity, we selected 8 PDX models (7 TNBC and 1 High Grade Serous Ovarian Carcinoma) showing different BRCA1 profiles. Our experimental PDX set comprised 2 BRCA1 wild type (BRCA1-WT) TNBC (b1995, b3804), 4 BRCA1-Me TNBC (b3977, b4122, 15b1516, 15b0018) and 2 (1 TNBC and 1 HGSOC) BRCA1 mutated (BRCA1-Mut) models (tm168, o10047) ( Table 1 ; Supplementary Figure 1 ). BRCA1-Me PDX models 15b1516, 15b0018 and b3977 were established from post-NACT residual tumors that had shown poor response to neoadjuvant treatment ( Table 1 ; Supplementary Figure 1 ). The two BRCA1-Mut models (tm168, o10047) and the remaining BRCA1-Me model (b4122) were established from treatment naïve tumors. Hypermethylation of the BRCA1 promoter region was determined in both the PDX models and the patient tumors using methylation specific PCR (MS-PCR). It is of note that no pre-treatment biopsy was available for the tumors that had undergone NACT, hence, BRCA1 promoter methylation was determined after NACT on these tumor samples. Of the four BRCA1-Me PDX, 3 presented homozygous (Me/Me) methylation (15b0018, b3977, b4122) and 1 hemizygous (Me/U) methylation (15b1516) ( Supplementary Figure 2).

PDX were grafted subcutaneously on Swiss-nude mice and olaparib was administered orally at 100 mg/kg, 5days per week for 5 weeks. In a parallel treatment group, mice received 50mg/kg Carboplatin (CBP) by intraperitoneal injection twice a week for 4 weeks, while the control group received daily the olaparib vehicle administered orally. Tumor volumes were measured twice a week. Three (tm168, b4122, o10047) of the 8 PDX showed disease stabilization or limited tumor size reduction under olaparib treatment, while the 5 other models (b3804, 15b0018, 15b1516, b1995, b3977) progressed ( Figures 1A, B ; Supplementary Figure 3A ). Of the 3 olaparib responders, PDX tm168 and o10047 were BRCA1-Mut and b4122 was BRCA1-Me, but established from a treatment naïve tumor. The 3 other BRCA1-Me models (15b0018, 15b1516 and b3977, which progressed under olaparib, were established from tumors that had received NACT. In the CBP arms we globally noted more favorable response patterns, with 6 of 8 PDX models showing partial to complete response and 2 PDX (b3804, 15b1516) progressing under treatment ( Figures 1A, C ; Supplementary Figure 3A ). Tumor growth was monitored after treatment end in 3 PDX, 2 BRCA1-Me TNBC (b3977, b4122) and 1 BRCA1-Mut HGSOC (o10047) ( Supplementary Figure 3B ) Both BRCA1-Me TNBC resumed growth shortly after end of olaparib administration. By contrast, the BRCA1-Mut o10047 exhibited complete regression 25 days after treatment ended and recurred after a lull of 3 weeks. In the CBP arm, progression after treatment end was observed in 1 of 3 PDX (b4122), while the two other models did not recur during the follow up period ( Supplementary Figure 3B ).

Targeted sequencing was performed on the 7 TNBC PDX to search for mutations affecting principal DNA damage response genes and targetable cancer genes. The HGSOC PDX was analyzed by exome sequencing as part of a previous study (55). Mutations were found in the TP53, PTEN, KRAS, RIF1 and STK11 genes (7, 2, 1, 1 and 1 PDX respectively) ( Supplementary Table 2 ). No mutations were detected in further DNA repair genes such as BRCA2, PALB2 or the RAD51 orthologs RAD51B, C or D. The 8 PDX models were also analyzed for copy number changes by array-CGH. Genetic instability and HRD scores were determined ( Figure 2A ). All TNBC models, irrespective of BRCA1 mutation or epigenetic silencing, presented elevated genetic instability scores (CX2, CX5) shown to be linked with BRCA-deficiency (20). We also noted that all tested TNBC models presented elevated HRD scores as defined by Abkevich and coworkers (39). Thus, elevated genetic instability and HRD scores suggested preexisting HR deficiency in the PDX models used in this study, including the two BRCA1-WT models.

We took advantage of the CGH analysis to determine copy number changes affecting key repair genes in the TNBC models. Except b3804, most PDX presented hemizygous losses affecting 3 to 4 HR genes of a list including BRCA1, BRCA2, PALB2, RAD51, RAD51B and RAD51C suggesting that the combined copy number reductions on these key HR genes could have contributed to a global HR attenuation and elevated genetic instability in these tumors ( Supplementary Figure 4 ). But we cannot exclude the existence of undetected genetic or epigenetic anomalies affecting genes involved in HR maintenance.

To determine the impact of promoter hypermethylation or coding mutations on BRCA1 protein expression, BRCA1 protein levels were measured by western blotting (WB) in the 8 PDX models. Noticeably, BRCA1-Me PDX b4122, b3977, 15b0018 and BRCA1-Mut o10047, tm168 showed no detectable BRCA1 band, suggesting a loss of BRCA1 functionality in these tumors. By contrast, the BRCA1-WT b1995, b3804 and the BRCA1-hemimethylated 15b1516 PDX presented detectable BRCA1 bands ( Figure 2B ). We noted that PDX b3804 expressed high levels of a short 100 kD BRCA1 protein isoform and low levels of the full length (268 kD) protein ( Figure 2B ). The 100 kD band is compatible with the size of the hypomorphic Δ11-BRCA1 isoform, which corresponds to a BRCA1 splice-variant where exon 11 is excluded and that is frequently expressed in tumors bearing mutations in exon 11 (56). However, no BRCA1 coding mutation, that could have explained the expression of the Δ11-BRCA1 isoform, was detected in PDX b3804 ( Supplementary Table 1 ). We also analyzed expression of the BRCA2, RAD51, PARP1 and 53BP1 proteins, which are important actors of HR. Whereas RAD51 expression appeared elevated and relatively constant in the different PDX models, that of BRCA2 was more variable, with no obvious link with BRCA1 status or olaparib response, except possibly the BRCA1-Mut PDX tm168, which showed very low BRCA2 protein levels ( Figure 2B ). PARP1 protein expression was detected in all PDX models, but, interestingly, the lowest levels were detected in the 3 BRCA1-Me (b3977, 150018 and 15b1516) that had been exposed to NACT and responded poorly to olaparib ( Figure 1A ). Finally, 53BP1 protein was expressed at high levels in 5 models and at low levels in 3 other PDX ( Figure 2B ).

In HR-proficient cells, BRCA1 and RAD51 proteins cluster at DNA damage sites forming nuclear foci that can be detected by immunofluorescence labeling. Absence of BRCA1 and/or RAD51 foci is considered as a sign of HR-deficiency. We scored the fraction of cells positive for BRCA1 and RAD51 foci in cryosections of PDX models sampled at the end of olaparib and vehicle treatment and searched for an association with olaparib response. Tumors presenting >10% cells showing >5 BRCA1 or RAD51 per nucleus were scored positive. PDX sections were immunolabeled with commercial antibodies directed against BRCA1, RAD51, and the S phase marker Geminin. Because of secondary antibodies species compatibility, BRCA1 and RAD51 were co-immunolabeled ( Figure 3A ; Supplementary Figure 5 ), whereas Geminin staining was performed separately to ascertain the presence of S phase cells in each sample. Immunolabeling scoring was performed in 3 independent PDX tumors per model on at least 100 cells per section. Geminin staining showed 10 to 30% of Geminin-positive cells, confirming that tissue section actually comprised S phase tumor cells ( Supplementary Figure 6 ). Four (4) of the 8 olaparib treated PDX, corresponding to 2 BRCA1-Mut (tm168, o10047) and 2 BRCA1-Me (b3977, b4122), scored negative for BRCA1-foci ( Figures 3B, J ), in agreement with the absence of BRCA1 band in the WB analysis ( Figure 2B ). BRCA1-foci positive models comprised two BRCA1-WT (b3804, b1995) and two BRCA1-Me PDX (b15b1516, 15b0018). Three of the four BRCA1-foci negative PDX (tm168, o10047, b4122) scored negative for RAD51-foci ( Figures 3A, C, J, K ) and corresponded to the 3 models that responded to olaparib ( Figures 1A, B ). Noticeably, PDX b3977, which progressed under olaparib ( Figures 1A, B ), scored positive for RAD51-foci, despite the fact it did not show BRCA1-foci ( Figures 3A, C, J, K ). All other RAD51-foci-positive PDX scored also positive for BRCA1 foci and showed mediocre response to olaparib ( Figures 1A, B ; 3A–C, J, K ). Noticeably, the 3 BRCA1-Me models established from NACT treated TNBC (15b0018, 15b1516, b3977) were RAD51-foci positive and showed scores similar to those observed in the 2 BRCA1-WT models (b1995, b3804) ( Figures 3B, J ). By contrast, the BRCA1-Me PDX b4122, which was not exposed to NACT prior PDX establishment, was both BRCA1 and RAD51-foci-negative and did not progress under olaparib. These data suggested that while b4122 was indeed HR deficient, the 3 NACT treated BRCA1-Me models were, at least partially, HR proficient.

We verified whether BRCA1 or RAD51 nuclear foci formation was associated with olaparib response ( Figures 3F-I ). We noted a significantly association of BRCA1 and RAD51 foci absence (or low levels) with olaparib response (t-test p=0.0017 and p<0.0001 respectively) ( Figures 3F, G ). Interestingly, the differences in foci positive cells between responsive and non-responsive cells were discernible in both tumor sections from olaparib treated ( Figures 3F, G ) and sections from vehicle treated tumors and showed equivalent statistical significance ( Figures 3J, K ).

We also scored γH2AX and 53BP1 foci, which are standard DNA damage markers, and observed that models presented 20 to 60% foci-positive cells for either marker, indicating severe DNA damage in olaparib treated PDX, to the exception of 15b1516 where only 5 to 10% cells scored positive ( Figures 3D, E ). Interestingly, we noted that 53BP1-foci tended to be more frequent in olaparib responsive PDX, compared with non-responsive PDX (t-test p=0.028) ( Figure 3H ). However, no difference was found with γH2AX-foci ( Figure 3I ).

It is well documented that BRCA-deficient tumors or cell lines rapidly acquire resistance to treatment associated with the restoration of HRR (56). We were, thus, interested to explore the restoration of RAD51-foci formation in BRCA1-deficient cell lines with acquired olaparib resistance. We, thus, isolated olaparib resistant clones from three cell lines; SUM159-KO1 and SUM159-KO2, two CRISPR BRCA1 knock out clones we engineered from the TNBC SUM159 cell line, the SUM149 TNBC cell line which bears a frameshift mutation in exon 11 (2288delT) (57) and the UWB1.289 ovarian cancer cell line also showing a frameshift mutation in exon 11 (2594delC) (58). Resistance was obtained by exposing cell cultures to incremental olaparib concentrations for at least 12 weeks. Olaparib resistant cell lines were designated with the suffix Re (SUM159-KO1-Re, SUM159-KO2-Re, SUM149-Re and UWB1.289-Re). We characterized BRCA1, BRCA2, RAD51 and PARP1 protein expression by WB, as well as RNA expression changes associated with the acquisition of olaparib resistance ( Figures 4A, B ). In SUM159-KO1 and SUM159-KO2 no BRCA1 protein was detected. SUM159-KO1-Re reexpressed the full length BRCA1 protein and SUM159-KO2-Re showed no difference in BRCA1 protein expression compared with it clone of origin. SUM149 and UWB1.289 expressed no full length BRCA1 and variable levels of the Δ11-BRCA1 100 kD band. Interestingly, SUM149-Re expressed the BRCA1 full length and increased levels of the Δ11-BRCA1 variant, whereas UWB1.289-Re showed strongly increased Δ11-BRCA1 levels ( Figure 4A ). No significant difference was found for BRCA2, RAD51 and PARP1 protein expression in the different cell line models. At the RNA expression level, we noted increased gene expression of RAD51 and RAD51C in SUM159-KO1, of RAD51C and ABCG2 in SUM149-Re, as well as of Δ11-BRCA1, RAD51C, RAD51D, PALB2 and ATM in UWB1.289-Re which could be related with their acquisition of olaparib resistance.

Next, we scored BRCA1, RAD51, 53BP1 and γH2AX foci formation in the olaparib treated cell lines ( Figure 5A ). Except in SUM159-KO2-Re which does not express BRCA1, nor ∆11-BRCA1, and scored BRCA1-foci-negative, BRCA1-foci scores tended to increase in olaparib resistant clones compared with cells or origin ( Figure 5B ). A similar trend was noticeable for RAD51-foci, whose numbers nearly doubled in olaparib resistant cells relative to cell lines of origin ( Figure 5C ). By contrast, 53BP1 and γH2AX foci numbers clearly decreased in olaparib resistant variants ( Figures 5D, E ). These differences in foci numbers between olaparib resistant cells and their original counterparts, increase for BRCA1 and RAD51, decrease for 53BP1 and γH2AX, were all statistically significant and in coherence with findings we made on PDX ( Figures 5F-I ). The reduction of γH2AX-foci associated with the increase of BRCA1 and RAD51-foci observed in resistant cells suggest a global decrease of DNA damage in these cells upon olaparib treatment due to restored HR capacity. We also derived olaparib resistant cells from the hemimethylated BRCA1-Me/UM HCC38 TNBC cell line, whose methylation status we confirmed. While BRCA1-foci scores only modestly increased in olaparib treated cells, the fraction of RAD51-foci positive cells doubled in HCC38-Re variant cells ( Supplementary Figure 7 ). Overall, our cell line data strongly support the notion that resistance to treatment and particularly to olaparib in BRCA1-deficient cancer cells is frequently associated with restoration of RAD51 foci formation, thus, signing for restored HR capacity.

As PDX responding to CBP represented twice the number of olaparib responders (6 vs. 3 respectively), we wanted to determine whether the response to CBP was associated with reduced RAD51 and/or BRCA1-foci formation ( Figures 1B, C) . Like in the olaparib arm, individual grafts that responded to CBP were predominantly BRCA1-foci-negative/RAD51-foci-negative ( Figures 6A, B ) and, interestingly, most grafts of the BRCA-foci-negative/RAD51-foci-positive model b3977, whose response to olaparib was mediocre, showed tumor size reduction under CBP. However, grafts of the BRCA1-positive/RAD51-positive PDX 15b0018 and b1995, respectively BRCA1-WT and BRCA1-Me and bad responders to olaparib, showed good response to CBP pointing to the fact that CBP efficacy is not solely based on the HR status, but can also rely on alternative DNA repair mechanisms ( Figure 6B ). Despite these two BRCA1-foci-positive/RAD51-foci-positive models, both BRCA1-foci-negativity and RAD51-foci-negativity were significantly associated (t-test p-value 0.0059 and 0.0176 respectively) with CBP response in our dataset ( Figure 6D ). Next, we computed the Sensitivity, Specificity and Accuracy of RAD51, BRCA1, 53BP1 and γH2AX foci in predicting the response to olaparib and CBP ( Table 2 ). Concerning olaparib response RAD51 foci showed high sensitivity (88%), specificity (82%) and accuracy (84%), while BRCA1 foci showed excellent sensitivity (94%) but lower specificity (66%) and accuracy (74%). Interestingly, when CBP response was considered, BRCA1 foci performed globally better showing 70% sensitivity, 92% Specificity and 75% accuracy in comparison with RAD51 foci, which showed 57% sensitivity, 100% Specificity but 68% accuracy. The performances of 53BP1 and γH2AX foci were overall rather contrasted showing excellent sensitivity (100%), but poor specificity (18%) and accuracy (41%) for olaparib response, while they reached better values for CBP response with high accuracy (87%) mitigating the mediocre specificity (46%). Overall, these data support the use of RAD51 foci for olaparib response prediction, while BRCA1 foci, possibly in combination with RAD51 foci, appear interesting for CBP response. While γH2AX and 53BP1 foci do not appear as convincing it may be interesting to reevaluate their performance for CBP response on a larger dataset.