Replication catastrophe is responsible for intrinsic PAR glycohydrolase inhibitor-sensitivity in patient-derived ovarian cancer models

Background Patients with ovarian cancer often present at advanced stage and, following initial treatment success, develop recurrent drug-resistant disease. PARP inhibitors (PARPi) are yielding unprecedented survival benefits for women with BRCA-deficient disease. However, options remain limited for disease that is platinum-resistant and/or has inherent or acquired PARPi-resistance. PARG, the PAR glycohydrolase that counterbalances PARP activity, is an emerging target with potential to selectively kill tumour cells harbouring oncogene-induced DNA replication and metabolic vulnerabilities. Clinical development of PARG inhibitors (PARGi) will however require predictive biomarkers, in turn requiring an understanding of their mode of action. Furthermore, differential sensitivity to PARPi is key for expanding treatment options available for patients. Methods A panel of 10 ovarian cancer cell lines and a living biobank of patient-derived ovarian cancer models (OCMs) were screened for PARGi-sensitivity using short- and long-term growth assays. PARGi-sensitivity was characterized using established markers for DNA replication stress, namely replication fibre asymmetry, RPA foci, KAP1 and Chk1 phosphorylation, and pan-nuclear γH2AX, indicating DNA replication catastrophe. Finally, gene expression in sensitive and resistant cells was also examined using NanoString or RNAseq. Results PARGi sensitivity was identified in both ovarian cancer cell lines and patient-derived OCMs, with sensitivity accompanied by markers of persistent replication stress, and a pre-mitotic cell cycle block. Moreover, DNA replication genes are down-regulated in PARGi-sensitive cell lines consistent with an inherent DNA replication vulnerability. However, DNA replication gene expression did not predict PARGi-sensitivity in OCMs. The subset of patient-derived OCMs that are sensitive to single-agent PARG inhibition, includes models that are PARPi- and/or platinum-resistant, indicating that PARG inhibitors may represent an alternative treatment strategy for women with otherwise limited therapeutic options. Conclusions We discover that a subset of ovarian cancers are intrinsically sensitive to pharmacological PARG blockade, including drug-resistant disease, underpinned by a common mechanism of replication catastrophe. We explore the use of a transcript-based biomarker, and provide insight into the design of future clinical trials of PARGi in patients with ovarian cancer. However, our results highlight the complexity of developing a predictive biomarker for PARGi sensitivity. Supplementary Information The online version contains supplementary material available at 10.1186/s13046-021-02124-0.


Background
Ovarian cancer is a leading cause of gynaecological cancer-related death, accounting for more than 200,000 deaths globally in 2020 [1,2]. The most prevalent subtype, high-grade serous ovarian cancer (HGSOC), accounts for approximately 70% of cases and is frequently lethal because it is highly aggressive and presents at an advanced stage where cure is unlikely [3]. Treatment options are limited, but typically include cytoreductive surgery plus paclitaxel/platinum chemotherapy, with maintenance therapy used in specific subgroups [4,5]. While many patients initially respond well, most develop recurrent disease, yielding 10-year survival rates of only ~ 35% [6][7][8][9].
HGSOC is characterised by ubiquitous TP53 mutations, and extensive copy number variation (CNV) [10,11], implying chromosome instability [12], and indeed, HGSOC is one of the most chromosomally unstable cancers [6,11,13]. Approximately 50% have homologous recombination (HR) deficiency (HRD), which in ~ 20% of cases is due to alteration of BRCA1 or BRCA2 [14]. While targeted therapies have led to personalised treatments across many cancers, this therapeutic strategy is limited in HGSOC due to the paucity of actionable driver mutations. Other therapeutic strategies are therefore required and indeed, HRD opened up an alternative: synthetic lethality. This approach was pioneered by the ability of poly(ADP-ribose) polymerase (PARP) inhibitors (PARPi) to selectively kill BRCA -mutant cells [15,16], and these drugs are now yielding major benefits for patients with HRD tumours [17][18][19][20]. However, this leaves ~ 50% of cases that are HR-proficient (HRP) and therefore unlikely to benefit from a PARPi. Furthermore, the majority of HRD tumours will eventually acquire PARPi resistance with treatment [21]. Thus, alternative strategies are required to improve outcomes for women with inherent and acquired resistance to both PARP inhibition and platinum-based chemotherapy.
To develop novel therapeutic strategies, attention is focussing on targeting poly(ADP-ribose) (PAR) glycohydrolase (PARG), the enzyme that counterbalances PARP1/2 activity [22][23][24][25][26][27]. Thus far, we have characterised PARG inhibitor PDD00017273, hereafter PARGi, a quinazolinedione that inhibits PARG with an in vitro IC 50 of 26 nM [28]. Upon analysing a panel of six ovarian cancer cell lines in response to PARGi and the PARP inhibitor Olaparib, hereafter PARPi, we discovered that OVSAHO, COV318, COV362 and CAOV3 proliferated in both inhibitors, while Kuramochi and OVCAR3 displayed differential sensitivities; Kuramochi proliferation was suppressed by PARGi, while OVCAR3 proliferation was suppressed by PARPi [26]. Sensitivity of Kuramochi to PARGi was accompanied by pan-nuclear γH2AX staining, which is indicative of replication catastrophe [29], and a synthetic lethal siRNA screen identified several DNA replication genes, including TIMELESS, that when inhibited sensitised OVCAR3 to PARGi. Furthermore, PARGi induced replication fork asymmetry in Kuramochi but not OVCAR3, suggesting that Kuramochi cells have an underlying DNA replication vulnerability that causes frequent fork stalling, and are thus more reliant on PARG to re-start stalled forks, possibly by reversing PARP-mediated inhibition of RECQ1 [26].
The notion that DNA replication vulnerabilities might confer PARGi sensitivity led to the identification of additional PARGi-sensitive cell lines; interrogation of Cancer Cell Line Encyclopedia (CCLE) data identified several ovarian cancer cell lines with down-regulated DNA replication genes, namely HS571T, RMG1, OV56, OVMANA and OVISE, as well as Kuramochi [26]. While HS571T is no longer available for research purposes, we obtained RMG1, OV56, OVMANA and OVISE, and showed that RMG1 and OVMANA are PARGi-sensitive. This raised the possibility that a DNA "replication stress" gene expression signature might have potential as a predictive biomarker of PARGi sensitivity. Importantly, like Kuramochi, RMG1 showed differential sensitivity to PARGi and PARPi. However, while OVMANA were more sensitive to PARGi than PARPi, the distinction was less clear-cut.
These observations indicate that PARG inhibitors may open up new opportunities to treat ovarian cancers. However, a number of questions remained unanswered. Firstly, is the PARGi sensitivity exhibited by different cell lines mediated via the same mechanism? While PARGi prevents proliferation of RMG1 and OVMANA, whether this was due to the replication catastrophe phenomenon exhibited by Kuramochi cells was not established. Secondly, does PARGi sensitivity correlate with downregulated DNA replication genes? While the CCLE data indicated that DNA replication genes are down-regulated in OV56 and OVISE, their PARGi sensitivity was unclear [26]. If resistant, the utility of a DNA "replication stress" gene expression signature as a predictive biomarker of PARGi sensitivity is uncertain. To address these two questions, we set out to perform a detailed analysis of a panel of ovarian cancer cell lines to determine (a) whether PARGi sensitivity is via a common replication catastrophe mechanism, and (b) whether PARGi sensitivity does indeed correlate with the relative expression levels of DNA replication genes. Moreover, to assess the translational opportunity of PARGi, we screened a panel of patient-derived ovarian cancer models (OCMs) [13] to determine whether any were PARGi sensitive and if this sensitivity is via a replication stress mechanism that also correlates with expression levels of DNA replication genes.

Ex vivo ovarian cancer models (OCMs)
Research samples were obtained with informed patient consent from the Manchester Cancer Research Centre (MCRC) Biobank. The MCRC Biobank is licensed by the Human Tissue Authority (license number: 30004) and is ethically approved as a research tissue bank by the South Manchester Research Ethics Committee (Ref: 18/ NW/0092). The role of the MCRC Biobank is to distribute samples and does not endorse studies performed or the interpretation of results. For more information, see https:// www. mcrc. manch ester. ac. uk/ resea rch/ mcrcbioba nk. Ex vivo ovarian cancer models (OCMs) were expanded from ascites samples from 32 patients, of which seven were published previously [13]. All patients were initially diagnosed with HGSOC, however four were re-classified to: cytological diagnosis of 'suspicion of adenocarcinoma arising from the gynaecological tract' (patients 87 and 195), moderately differentiated (intermediate grade/grade 2) serous adenocarcinoma (patient 152), and low-grade serous ovarian cancer (patient 64) [30]. The age at diagnosis ranged from 44 to 84 years, the mean age at diagnosis was 62.8 years. Five samples were chemonaïve (patients 87, 110, 99, 195 and 231).

Colony formation assay
For established ovarian cancer cell lines, 1000 cells per well were seeded into 6-well plates 24 h prior to drug treatment. For OCMs, 2000 cells per well were seeded into either Primaria or collagen coated (50 μg/mL of BD collagen type I, rat tail, [BD Biosciences Discovery Labware] in 0.02 N Acetic Acid for 1 h prior to seeding) 6-well plates 24 h prior to drug treatment. For OCM.165, tumour cells were seeded at 4000 cells per well 24 h prior to drug treatment. DMSO (control), PARGi, PARGi-Me or PARPi were added; cells were treated either continuously or the drugs were washed out at the specific time points as indicated in the figures. Following colony formation (range: 2-4 weeks for ovarian cancer cell lines; range: 2-12 weeks for OCMs), colonies were fixed in 1% formaldehyde for 10 min, stained with 0.05% (w/v) crystal violet solution (Sigma-Aldrich) for 10 min and rinsed with ddH 2 O. Plates were imaged using a ChemiDoc ™ Touch Imaging System (BioRad), and analysed with an ImageJ 'colony area' plug-in [32].
For high-throughput immunofluorescence, cells were processed as above in 96-well plates (PerkinElmer Cell Carrier) with two additional final washes in PBS. For RMG1, wells were coated with 0.01% Poly-L-Lysine. For OCMs, plates were collagen coated (as above) prior to plating. For ovarian cell lines, cells were seeded at 2000-28,000 per well. For OCMs, cells were seeded at 750-6000 cells per well. Images were acquired using Operetta ® High Content Imaging System (PerkinElmer) and quantified using Columbus High Content Imaging and Analysis Software (PerkinElmer). Mean fluorescence intensity or foci quantification within the nuclear area (demarcated using Hoechst stain) using Columbus 'spot finder' tool, was quantified as a mean value per cell. These were also calculated for secondary antibody-only exposed cells (control). For final values, secondary antibody-only control mean values were subtracted from the stained cell mean values.
The 12-well plates were centrifuged at 300×g for 2.5 h at 30 °C. 1 mL of culture media was added and the plates incubated overnight. Puromycin (Sigma Aldrich) (2 μg/ ml for OVMANA, 1 μg/ml for all other ovarian cancer cell lines and OCMs) was added 48 h after transduction.

Drug sensitivity assay and cell fate profiling
Ovarian cancer cell lines were seeded, at 500-8000 cells per well and OCMs at 750-6000 cells per well, into a 96-well plate (Greiner Bio-One) 24 h prior to drug treatment. For OCMs, 96-well plates were coated with collagen (as above) prior to plating. To determine half maximal effective concentration (EC 50 ), cells expressing GFP-H2B were used and PARGi, PARGi-Me, PARPi, niraparib tosylate and cisplatin were serially diluted from 100 μM-0.381 nM (19 concentrations in total). Paclitaxel was serially diluted from 10 μM-0.0381 nM (19 concentrations in total). Following drug treatment, cells were imaged using an IncuCyte ® ZOOM (Essen BioScience) equipped with a 20X objective and maintained at 37 °C in a humidified 5% CO 2 atmosphere for ovarian cell lines or a humidified 5% CO 2 and 5% O 2 atmosphere for OCMs. Nine phase contrast and fluorescence images (for GFP-H2B expressing cells) were collected per well every 4 or 6 h for 120 h to analyse cell proliferation. To determine cell fates, cells were treated with DMSO, PARGi or PARPi, and cells were imaged every 10 min for 120 h.
IncuCyte ® ZOOM software was used in real-time to measure green object count in cells expressing GFP-H2B, as a proxy for cell proliferation. Green object count was used to generate dose-response curves in Prism (Graph-Pad) from which EC 50 values were calculated. In highly resistant cells, in which EC 50 could not be determined accurately, the EC 50 assigned a value of 50 μM (i.e. half the maximum dose tested). To generate cell fate profiles, image sequences were exported in MPEG-4 format and analysed manually to time and annotate cell behaviours [34]. Prism 8 (GraphPad) was used for statistical analysis and presentation.

Immunoblotting
Ovarian cancer cells lines were treated with DMSO, PARGi or PARPi for 48 h prior to harvesting, or for 2 h with 2 mM Hydroxyurea as a positive control. OCMs were treated with DMSO, PARGi or PARPi for 96 h prior to harvesting, or exposed to 10 Gy of x-ray ionising radiation (IR) using the Faxitron ® (Hologic), 4 h prior to harvesting, as a positive control. Proteins were extracted, quantified by Bradford assay, then denatured by boiling in sample buffer (0.35 M Tris pH 6.8, 0.1 g/ml sodium dodecyl sulphate, 93 mg/ml dithiothreitol, 30% glycerol, 50 μg/ml bromophenol blue). Proteins were resolved by SDS-PAGE and electroblotted onto Immobilon-Fl PVDF membrane (Millipore; LI-COR) or Immobilon -P Transfer Membrane (Millipore). For LI-COR imaging, in place of a loading control REVERT total protein stain solution (LI-COR) was used for normalisation: membrane was incubated with REVERT solution for 5 min, followed by washing in 6.7% (v/v) glacial acetic acid in water, 30% (v/v) methanol in water. Before imaging on the Odyssey ® CLx Imaging System (Li-COR), the membrane was washed with water. Following imaging, REVERT stain was removed using REVERT reversal solution (0.1 M NaOH, 30% v/v methanol in water). Membranes were then blocked using 5% dried skimmed milk or 5% BSA (for anti-pChk1) diluted in TBS-T (50 mM Tris pH 7.6, 150 mM NaCl, 0.1% Tween-20).

Functional Rad51 assay
OCM.109, OCM.246, OVCAR3 and Kuramochi cells were seeded overnight onto collagen coated (see above) 19 mm coverslips at a cell density of 112,000, 56,000, 36,000 and 36,000 cells per coverslip, respectively. The following day, cells were exposed to 2 Gy of X-ray IR and 1 μM of PARPi or DMSO control for 24 h. Cells were stained for CENPF and Rad51 by the previously stated immunofluorescence protocol. To determine the HR status, the number of CENPF-positive nuclei containing ≥5 Rad51 foci (CENPF+Rad51+) was calculated for 10 separate fields of view, using the Axioskop 2 (Zeiss) microscope, at 40x magnification. A < 2-fold rise in the ratio of Rad51 + CENPF+ cells/CENPF+ cells following IR plus PARPi treatment versus control, was reported as HRD.

DNA fibre assay Sample preparation
Sub-confluent cells were incubated in the presence of DMSO or PARGi for 48 h, then pulsed with BrdU at 5 μM plus DMSO or PARGi for 20 min. This was followed by 3 washes with warm PBS, pulsing with 200 μM IdU (Sigma-Aldrich) plus DMSO or PARGi for a further 20 min, then washing twice with ice-cold PBS. Following trypsinisation using 0.05% Trypsin-EDTA (Gibco), cells were diluted in ice-cold PBS to give a final concentration of 1-5 × 10 5 cells/ml and kept on ice.

Slide preparation
The cell suspension (2 μl) was then dropped onto microscope slides and dried at room temperature for 5-10 min before mixing with 7 μl of spreading buffer (200 mM Tris-HCl pH 7.5, 50 mM EDTA, and 0.5% SDS), and incubating for a further 5 min. Slides were tilted approximately 5-10° so that the cell suspension runs across the length of the slide. Slides were air dried and fixed in methanol/ acetic acid (3:1) for 10 min, air dried and stored at 4 °C.

Imaging and quantitation
Images were acquired using an Axioskop 2 (Zeiss) microscope fitted with a CoolSNAP HQ camera (Photometrics) and 2-5 slides analysed per condition. Fibre lengths were quantified using ImageJ software (NIH).

BRCA1/2 genotyping by next generation sequencing (NGS)
DNA from OCMs was extracted using the cobas ® DNA Sample Preparation Kit (Roche). Library enrichment used the GeneRead DNAseq BRCA1 & BRCA2 version 2 kit (Qiagen). 20 ng DNA was amplified in 4 multiplex primer pools. Following PCR-based target enrichment, library preparation and purification followed a custom protocol using AMPure XP beads (Beckman Coulter) for size selection and TruSeq PCR-Free indexes and reagents for indexing (Illumina). The DNA library was then paired-end sequenced on an MiSeq (Illumina) with v2 chemistry (2 × 150 cycles). Bioinformatic analysis used an in-house pipeline, which has been validated to detect low level mosaic calls down to a variant allele fraction of 4% and used a software consensus between VarScan v2.3.6 and DREEP v0.7. Large indel events were assessed using Pindel v0.2.4.t. The NGS assay was able to detect singlenucleotide variants and duplications, deletions and insertions ≤40 base pairs in length, across the whole coding sequence of BRCA1 and BRCA2 +/− 15 base pairs beyond each exon-intron junction. The target read depth across all coding exons was a minimum of 350X. All variant calls identified following bioinformatics analysis were independently reviewed within a genome browser. At a variant allele frequency ≥ 4% had a call sensitivity > 95% and specificity > 99% after a manual review.

BRCA1/2 genotyping by multiplex ligation probe amplification (MLPA)
Testing for genomic rearrangements/copy number variation in BRCA1/2 was performed by MLPA using the MRC Holland probe kits P002-D1 (BRCA1) and P045-C1 (BRCA2). Amplified ligation products were subject to fragment analysis using an ABI 3130xl Genetic Analyser and size called using GeneMapper v2.0 (Applied Biosystems). Copy number status calling was performed using data exported from GeneMapper using custom-developed MLPA spreadsheets that report relative dosage quotient for each probe compared with reference control samples. All MLPA analysis assays were performed in duplicate for confirmation of results.

NanoString of ovarian cancer cell lines
RNA was extracted using RNeasy Plus Mini kit (Qiagen) as per manufacturer's instructions. 100 ng RNA was provided at concentration of 4 ng/μl to the Genomic Technologies Core Facility at University of Manchester for analysis using the NanoString nCounter Analysis System. Twenty-five replication genes, 25 mitotic genes, 25 apoptotic genes (Fig. 4B) and 4 reference genes as normalisation controls appropriate for use in ovarian cancer (RPLP0, PPIA, IPO8, TBP) were analysed using nSolver ™ Analysis Software 4.0., which generated normalised transcript counts.

Analysis of mitotic, DNA replication and apoptotic gene expression in RNAseq datasets from the CCLE and Klijn et al., (2015)
Previous analysis utilised gene expression microarray profiling from the CCLE to identify ovarian cancer cell lines with low expression of DNA replication genes [26,38]. Here, we updated this analysis to utilise the more recent RNAseq performed by the CCLE [39] (Fig. 4C). Briefly, read counts were obtained from the Broad Institute data portal (https:// porta ls. broad insti tute. org/ ccle). The R package DESeq2 (v1.26.0) was used to normalise and apply a variance stabilising transformation to the assembled read count matrix, followed by z-score transformation.
We additionally utilised read counts for the panel of 10 ovarian cancer cell lines of the same DNA replication, mitotic and apoptotic gene lists from a pan-cancer cell line RNAseq study [40] (GEO accession number GSE40788). Raw read counts were treated as described previously. This revealed higher expression of these genes in OV56 cells than indicated in the CCLE data set, in agreement with our NanoString data (Fig. S4A).

RNA-sequencing and analysis
RNA was extracted using RNeasy Plus Mini kit (Qiagen), quantified using a Qubit fluorometer (Life Technologies) and quality/integrity assessed using a 2200 TapeStation (Agilent Technologies). Sequencing libraries were then generated using the TruSeq ® Stranded mRNA assay (Illumina, Inc.) according to the manufacturer's protocol. Adapter indices were used to multiplex libraries, which were pooled and paired-end sequenced on an Illumina HiSeq4000 instrument. The output data was demultiplexed (allowing one mismatch) and BCL-to-Fastq conversion performed using Illumina's bcl2fastq software.
Overlapping paired reads were merged using BBMerge, and trimming and filtering was done using BBDuk, both from BBMap v36.32. The filtered reads were mapped to the human reference sequence analysis set (hg38/Dec. 2013/GRCh38) from the UCSC browser, using STAR v2.7.2b [41]. The genome index was created using the comprehensive Gencode v32 gene annotation. The number of reads per gene were counted using '--quantMode GeneCounts' within the STAR command.

Identification of PARGi sensitive and resistant ovarian cancer cell lines
To analyse the molecular mechanisms underlying PARGi sensitivity, we assembled a panel of ten ovarian cancer cell lines: OVMANA, Kuramochi, OVISE, RMG1, OV56, CAOV3, COV362, OVCAR3, OVSAHO and COV318 (Fig. 1A). While six of these cell lines have genetic hallmarks of HGSOC, in particular TP53 mutations (Fig. 1A), OVMANA, OVISE, RMG1 and OV56 are more representative of ovarian clear cell carcinoma [30,45]. Nevertheless, we reasoned that analysing a panel of epithelial ovarian cancer cell lines with potential for PARGi sensitivity would provide insight into the intrinsic vulnerabilities responsible for PARGi sensitivity, and in particular provide insight into whether PARGi sensitivity was occurring via a common mechanism.
Previously, we showed that Kuramochi, OVMANA and RMG1 are PARGi sensitive in long-term colony formation assays, while COV362, CAOV3, OVCAR3, OVSAHO and COV318 were resistant in a shortterm proliferation assay. While OV56 was identified as resistant by colony formation assay, OVISE was also classified as resistant because PARGi treatment failed to induce pan-nuclear γH2AX [26]. Note however that in our previous study, analysis of OVISE was challenging due to technical issues, which we have now overcome. To systematically compare PARGi sensitivity, all ten cell lines were now analysed in long-term colony formation assays following exposure to 1 μM PARGi for 24, 48, 72 h or continuously, and colony area quantitated. Examination of continuously treated cells confirmed that Kuramochi, OVMANA and RMG1 are indeed PARGi sensitive (Fig.  S1A, B). However, in contrast to our previous report, this latest analysis shows that OVISE is also PARGi sensitive. Outgrowth of OVMANA, Kuramochi and OVISE was also inhibited following drug exposure for 48 and 72 h, while RMG1 was not (Fig. S1C). In line with our previous study, growth of OV56, CAOV3, OVCAR3, OVSAHO and COV318 was not markedly suppressed even in the continuous presence of PARGi, while COV362 growth was partially suppressed (Fig. S1B). In addition, two out of four PARGi-sensitive cell lines were resistant to PARPi (Kuramochi and RMG1), while four out of six PARGiresistant lines were sensitive to PARPi (Fig. S1A, G, H), confirming differential sensitivity in some cases, despite both inhibitors targeting PARylation.
To independently analyse PARGi sensitivity, each cell line harbouring a GFP-tagged histone was analysed by time-lapse microscopy, over a 120-h period, in the presence of increasing concentrations of PARGi (Fig. S1D). Proliferation curves were then analysed to determine nuclear doubling rates and in turn calculate EC 50 values (Fig. S1E). COV318, OVSAHO, OVCAR3, COV362 and CAOV3 were largely unaffected even at very high concentrations of PARGi, yielding high or indeterminant EC 50 values (Fig. S1F). By contrast proliferation of RMG1, OVMANA and OVISE was strongly inhibited, yielding EC 50 values in the micromolar range (Fig. S1F). Consistent with our previous analysis, Kuramochi were particularly sensitive with an EC 50 value of 46 nM.
To integrate these two data sets, we plotted EC 50 values from the proliferation assays against colony area in the continuous presence of PARGi (Fig. 1B). This clearly highlights OVMANA, Kuramochi, OVISE and RMG1 as PARGi-sensitive, and OVSAHO, CAOV3, OVCAR3 and COV318 as PARGi-resistant. COV362 and OV56 are more ambiguous. While OV56 yields an EC 50 of ~ 4.5 μM, these cells clearly form colonies in 1 μM of PARGi and closer inspection reveals that these colonies are less dense than controls (Fig. S1A), suggesting that although PARGi slows OV56 proliferation it does not block it enough to prevent outgrowth. With an indeterminant EC 50 , COV362 appears PARGi-resistant, and closer inspection reveals large colonies indicating outgrowth ( Fig. S1A). Thus, we conclude that both COV362 and OV56 are PARGi resistant. In summary, taking together the long-term colony formation assay and the short-term nuclear proliferation assay, we conclude that OVMANA, Kuramochi, OVISE and RMG1 are PARGi-sensitive while OV56, CAOV3, COV362, OVCAR3, OVSAHO and COV318 are PARGi-resistant.

PARGi stabilises PAR chains in both sensitive and resistant cell lines
A possible explanation for PARGi resistance could simply be that the PARGi fails to engage with its target in resistant cells, for example due to drug efflux mechanisms. Therefore, we sought to determine whether PARGi was indeed inhibiting PARG activity in all ten cell lines, by measuring the accumulation of PAR chains. Analysis of single cells by immunofluorescence microscopy showed that in all cell lines, exposure to PARGi increased the intensity of nuclear PAR staining ( Fig. S2A-C). Importantly, increased PAR staining was blocked by cotreatment with an equimolar concentration of PARPi, indicating that accumulation of PAR chains was dependent on PARP activity ( Fig. S2A-C). Immunoblotting also revealed increased PAR levels in PARGi-treated versus untreated cells, with accumulation of higher molecular weight PAR species indicating stabilisation of PAR chains ( Fig. S2D-F). However, the very high molecular weight PAR species did not accumulate in OV56 or Kuramochi, and indeed, there is inter-line variation in PAR-chain dynamics (Fig. S2E). For OV56, although increased PAR staining with PARGi treatment was seen by microscopy, the increase was minimal by immunoblotting. This disparity is perhaps due to differences in assay sensitivity and/or arises from observing individual cells versus the whole population. Nevertheless, the two approaches support that PARGi resistance is not explained by a failure to stabilise PAR chains, but rather by an ability to tolerate the presence of stabilised PAR chains.

PARGi sensitivity is accompanied by markers of replication stress
Having established that PAR chains accumulate with PARGi treatment in all ten cell lines, we asked whether PARGi sensitivity occurs via a similar mechanism, in particular via DNA replication catastrophe. Previously, we showed that PARGi-sensitive Kuramochi cells displayed features consistent with persistent replication stress, DNA damage and eventually replication catastrophe upon prolonged exposure to PARGi [26]. If this phenomenon is a common cause of PARGi sensitivity, we reasoned that these features should also manifest in other sensitive lines, but not resistant lines. To test this, we measured accumulation of three well recognised markers of replication stress, namely γH2AX foci, RPA foci, and nuclear phospho-KAP1 [29,46,47] using highthroughput immunofluorescence microscopy ( Fig. 2A, B and S3A, B). After 48 h of exposure to PARGi, γH2AX foci were substantially and significantly increased in the sensitive lines OVISE, Kuramochi and OVMANA, but not in the six resistant cell lines ( Fig. 2A and S3A, B). By 72 h, cells with pan-nuclear γH2AX staining also became apparent ( Fig. 2A, B). Induction of RPA1 foci and nuclear phospho-KAP1 showed a similar trend ( Fig. 2A, B and  S3A), confirming a correlation between PARGi sensitivity and markers of replication stress. To independently analyse replication stress, we analysed cell populations by immunoblotting to measure phosphorylation of Chk1, normalised to total Chk1, and using the ribonucleotide reductase inhibitor hydroxyurea (HU) as a positive control (Fig. 2C). This was initially done by chemiluminescent western blot (Fig. S3C), but subsequently repeated using LI-COR system to enable quantification (Fig. 2C,  S3D). Because Chk1 phosphorylation in response to HU exhibited substantial variation between the cell lines, we expressed the phosho-Chk1/total-Chk1 ratio as a percentage of the HU-induced maximum. This shows that PARGi induced strong Chk1 phosphorylation responses in three of the sensitive lines RMG1, OVISE and Kuramochi, but not in the resistant lines (Fig. S3C, D), further suggesting a correlation between PARGi sensitivity and replication stress.
We did notice a few exceptions to this overall trend. While RMG1 cells are sensitive and showed a substantial fold change in RPA foci and strongly induced phosho-Chk1, they showed only modest increases in γH2AX foci and phosphorylated-KAP1 (Fig. S3A, D). Also, while CAOV3 cells are PARGi resistant, they showed induction of RPA1 foci, but not the other markers, suggesting this level of replication stress is tolerated sufficiently to allow growth (Fig. S3A). Indeed, it has previously been reported that PARG is required for recovery from persistent, but not short-term replication stress [48]. These exceptions highlight the limitations of relying on a single marker of replication stress. Therefore, we integrated these four datasets by plotting the fold changes of γH2AX foci versus RPA foci, phospho-KAP1 and phospho-Chk1 (Fig. 2D). This analysis shows a clear demarcation of sensitive and resistant cell lines, with OVISE, Kuramochi, RMG1 and OVMANA segregated from the six resistant cell lines. Thus, in toto, these observations demonstrate that PARGi sensitivity does reflect a common mechanism in ovarian cancer cell lines, namely persistent replication stress leading to replication catastrophe.

PARGi induces replication fork asymmetry in sensitive cell lines
We previously showed that inhibiting PARG causes DNA replication fork asymmetry in sensitive Kuramochi cells, but not resistant OVCAR3 cells [26], reflecting the role of PARG in restarting stalled replication forks. Having established that all four sensitive lines demonstrated features of replication stress when exposed to PARGi, we asked whether this was accompanied by persistent fork stalling, which we assessed using fork asymmetry assays [49]. Following a 48-h exposure to PARGi, cells were pulsed with BrdU to label active DNA replication forks, then chased with IdU to measure the speed   (Fig. 1A); accordingly, fork asymmetry following PARGi-treatment could be due to an absence of both PARG-and HR-mediated fork restart [27]. However, PARGi-resistant OVS-AHO has a homologous deletion in BRCA2 and does not show increased fork asymmetry following treatment. Furthermore, asymmetry only increased marginally in PARGi-sensitive OVISE, which is BRCA1/2 wild-type, although these cells have intrinsically high levels of asymmetry, possibly indicating an existing fork-restart vulnerability in these cells (Figs. 1A and 2G). The level of fork asymmetry conferring PARGi-sensitivity may therefore differ between cell lines, and is likely modulated by other factors in addition to HR status, such as the ability to utilise dormant replication origins. Indeed, there is variation in basal level of fork asymmetry across the panel (Fig. 2G). Nevertheless, we conclude therefore that PARGi sensitivity generally is accompanied by an induction of DNA replication fork asymmetry, indicating persistent fork stalling, further supporting the notion that PARGi sensitivity is due to persistent replication stress.

PARGi suppresses mitotic entry in sensitive cell lines
Persistent replication stress activates intra-S and G 2 /M checkpoints, thereby blocking cell cycle progression, and in particular inhibiting entry into mitosis. Indeed, we previously showed that a substantial fraction of PARGitreated Kuramochi cells underwent a Wee1-dependent G 2 arrest, and that those cells that did progress through mitosis often did so abnormally, and subsequently died [26]. To determine whether a similar cell fate was shared by other PARGi-sensitive lines, we analysed the panel of ten cell lines using time-lapse microscopy for 115 h and generated cell fate profiles [34]. In the absence of PARGi, the vast majority of cells in each line underwent multiple successful divisions and were alive at the end of the experiment (Fig. 3A). Significantly, PARGi treatment of the sensitive cell lines resulted in an increased proportion of cells that did not enter mitosis, with 48, 32 and 51% of OVMANA, Kuramochi and RMG1, respectively, blocking in interphase (Fig. 3A). This phenotype was not observed in the resistant cell lines where the majority of cells continued to complete multiple cell divisions when treated with PARGi.
Interestingly, in the four sensitive cell lines, a small proportion of cells did not enter mitosis in the absence of PARGi, consistent with these cells exhibiting an underlying vulnerability that is markedly exacerbated by inhibition of PARG. Of the resistant lines, a small proportion of COV318 cells also did not enter mitosis in the absence of PARGi, but in contrast to the sensitive lines, this did not increase substantially upon exposure to PARGi.
In most of the cultures analysed, a fraction of cells underwent abnormal mitoses and/or cell death and this phenotype was exacerbated by PARGi in sensitive but not resistant lines. In particular, while only 14% of PARGitreated OVISE cells blocked in interphase, the number of cells undergoing abnormal mitoses and/or cell death increased from 17 to 46% (Fig. 3A). Although the analyses above indicate that OV56 is PARGi-resistant, the cell fate profiling indicates that drug exposure increases the number of abnormal mitoses from 23 to 35% (Fig. 3A), possibly reflecting a synthetic effect with the prolonged cell culture and/or imaging conditions.
To quantitate these PARGi effects on mitotic potential, we counted the number of productive cell divisions in control and drug-treated populations. In sensitive cell lines, PARGi dramatically reduced the number of normal mitoses, by an average of 73%, compared with a minor reduction (11% on average) in resistant lines (Fig. 3B). Thus, considering both the cell fate profiles and the quantitation of successful divisions, we conclude that sensitivity to PARGi is accompanied by a dramatic loss of mitotic potential, in particular a pre-mitotic block, consistent with activation of intra-S and G 2 /M checkpoints due to persistent replication stress.

PARGi sensitivity correlates with replication stress
The analyses described above show that markers of replication stress ( Fig. 2A-D), fork asymmetry ( Fig. 2E-F) and a pre-mitotic block (Fig. 3), associated with increased nuclear size and reduced cell count (data not shown), also appear to correlate with PARGi sensitivity. To confirm this, we integrated the various datasets by scaling each parameter so that the most resistant and sensitive lines scored 1 and 0 respectively (Fig. 4A). These were then averaged to yield an overall sensitivity score and rank ordered. Importantly, the various cell biological parameters clearly align with sensitivity as defined by the colony formation and short-term proliferation assays, with COV318, OVSAHO, OVCAR3 and COV362 ranking as the most resistant lines, and OVMANA, Kuramochi and OVISE the most sensitive. Of the sensitive lines, RMG1 ranks as the least sensitive and indeed, as noted above, this line requires continuous exposure for fully penetrant Assay results (except EC 50 values) were scaled from 0 to 1, averaged, and cell lines rank ordered accordingly. γH2AX at 48 h is based on nuclear foci quantification. γH2AX at 72 h is based on nuclear intensity. B NanoString DNA replication, mitotic and apoptotic gene Z-scores of cell lines represented by heatmaps where red = high expression; white = average expression and blue = low expression. C Box-whisker plots (min-max) for DNA replication genes Z-scores shown for NanoString compared with RNA-sequencing (RNAseq) data from the CCLE [39]. See also Fig. S4 inhibition ( Fig. S1A-C). Of the resistant lines, OV56 is the least resistant and, again as noted above, although colonies form in the presence of PARGi, they are less dense than control colonies (Fig. S1A, B). CAOV3 are also at the more 'sensitive' end of the resistant spectrum and interestingly, these cells do display increased RPA foci when exposed to PARGi (Fig. S3A), suggesting that while PARGi does induce replication stress, it is resolved sufficiently to allow efficient recovery. Nevertheless, this integrated analysis confirms that PARGi sensitivity does indeed correlate with markers of replication stress in ovarian cancer cell lines, suggesting a common underlying mechanism. Moreover, because PAR chains are stabilised in all cell lines analysed (Fig. S2C, D), an important corollary is that resistant cell lines can efficiently complete DNA replication despite the presence of stabilised PAR chains.

PARGi sensitivity correlates with lower expression of DNA replication genes in cell lines
The strong correlation between markers of replication stress and PARGi sensitivity supports our previous conclusion that sensitive cell lines harbour an underlying DNA replication vulnerability that makes them particularly dependent on PARG activity to re-start stalled replication forks [26]. While the nature of this vulnerability remains to be determined, when we previously interrogated CCLE microarray data, we noted that in addition to Kuramochi, the expression levels of a number of DNA replication genes was lower in RMG1, OV56, OVMANA and OVISE cell lines compared with resistant lines [26]. Of these four, we previously showed that RMG1 and OVMANA are PARGi-sensitive, and here we demonstrate that OVISE is also sensitive. However, while OV56 is the most 'sensitive' of the resistant lines, it is resistant despite apparent low expression levels of DNA replication genes. To address this anomaly, we set out to independently validate the expression levels of 25 DNA replication genes using a custom NanoString Code-Set (Fig. 4B) and compare the overall expression of these genes with more recent RNA-sequencing (RNAseq) from the CCLE project [39] (Fig. 4C). NanoString analysis confirmed that expression of the 25 DNA replication genes is lower in PARGi-sensitive than -resistant lines, and for eight cell lines the CCLE and NanoString data significantly correlated (Fig. S4A). While the correlation between NanoString and CCLE for RMG1 is not significant, both find that this PARGi-sensitive line has lower expression of these genes than resistant lines. However, in contrast with the CCLE data, we find that OV56 have relatively high expression of the DNA replication genes in line with other resistant lines (Fig. 4C, S4A). Interestingly, while the more recent RNAseq data from the CCLE project is consistent with their previous microarray data [38,39], an independent RNAseq study of cancer cell lines also indicates higher expression of DNA replication genes in OV56 [40]. Indeed, the OV56 data from Klijn et al., significantly correlated with our NanoString expression data, but not the CCLE data (Fig. S4A). Aside from the anomaly of OV56, two independent RNAseq datasets [39,40] and our NanoString-based analysis therefore indicate that cell lines that are sensitive to PARGi tend to have lower expression levels of a number of DNA replication genes. If PARGi-sensitivity specifically correlates with DNA replication gene expression, we expected the expression level of genes involved in other processes not to be associated with sensitivity. To test this, we expanded our analysis to include 25 mitotic genes and 25 apoptotic genes, based on NanoString CodeSets we used previously [50]. While apoptotic gene expression was variable amongst both sensitive and resistant cell lines, expression of the mitotic genes was lower in PARGisensitive versus -resistant lines (Fig. 4B, S4A). However, we also observed a striking correlation between expression of replication and mitotic genes in our cell line panel (NanoString: R 2 = 0.87, p < 0.0001 [not shown]; CCLE: R 2 = 0.96, p < 0.0001) (Fig. S4B). Replication and mitotic gene expression also correlated in a wider cohort of 747 cancer cell lines within the CCLE database (R 2 = 0.72, p < 0.0001), but this was not observed for replication versus apoptotic genes (R 2 = 0.081) (Fig. S4B). Therefore, while the NanoString analysis builds on our previous analysis [26], showing association between low expression of DNA replication genes and PARGi sensitivity, we cannot rule out a wider cell cycle control gene expression phenotype.

Differential sensitivity to PARPi and PARGi
As PARG acts as a direct counterbalance to PARP-1/2, and further therapeutic strategies are needed for disease with inherent or acquired PARPi resistance, we wondered whether PARGi-sensitive cell lines showed differential sensitivity to PARPi, and vice versa. In terms of PARPi sensitivity, using a short-term confluence assay, we previously showed that OVSAHO, COV318, COV362 and CAOV3 were PARPi-resistant, whereas OVCAR3 was sensitive [26]. Because long-term target engagement can be required to observe synthetic lethality between cells with a BRCA1/2 mutation and Olaparib [28], we revisited this and compared PARPi and PARGi sensitivity in long-term colony formation assays with continuous drug exposure. Of these four lines, with the exception of COV318, all showed a reduction in outgrowth in response to PARPi (Fig. S1G). By contrast, and in agreement with our previous report, OV56, Kuramochi and RMG1 were relatively resistant to PARPi. Notably, OVISE and OVMANA were sensitive to both PARPi and PARGi. Thus, we conclude that of the ten ovarian cancer cell lines analysed here, four are PARPi-sensitive and PARGi-resistant; two are PARGi-sensitive and PARPiresistant; two are sensitive to both treatments and two are resistant to both (Fig. S1H). This implies that sensitivity to PARGi and PARPi is neither mutually exclusive nor overlapping, indicating that the determents of sensitivity are context dependent. In this regard OV56 is an interesting case -while categorised as PARGi-resistant, it is the least 'resistant' of the resistant lines and PARGi exposure is clearly not completely inconsequential: the shortterm proliferation assay yields high EC 50 values for PARPi and PARGi (Fig. S1F, G), and the colonies are less dense than controls for the PARGi-treated culture (Fig. S1A). In contrast, in response to PARPi, it forms robust colonies (Fig. S1A). Thus, despite the intimate relationship between PARP and PARG activities, and despite the complexities described above, a differential sensitivity clearly manifests across multiple cell lines, indicating that PARG inhibitors may offer an alternative therapeutic option to target a subset of ovarian cancers, such as HRP tumours that are less likely to respond to PARPi.

Identification of PARGi-sensitive patient-derived ovarian cancer models
Having demonstrated that PARGi sensitivity in established ovarian cancer cell lines arises due to replication catastrophe, we determined the impact of PARGi on a panel of 32 patient-derived OCMs from a living biobank focused on women treated at The Christie Hospital [13]. In contrast to established cell lines, these OCMs are clinically annotated and available for analysis at low passage. Seven OCMs were described previously [13], and with 25 additional OCMs described here (Supplementary  Table 1). Importantly, these new OCMs also displayed the features of HGSOC, namely nuclear atypia, universal expression of CK7 and PAX8, aberrant p53 expression, TP53 and BRCA1/2 mutations, and extensive chromosomal instability (Fig. S5; Supplementary Table 1). To screen for PARGi sensitivity, the panel was exposed to 1 μM drug for 96 h and colony area (ca) measured after ≥2 weeks (Fig. S6A). Of the 32, 15 were largely unaffected (ca ≥ 90%), 10 partially affected (50% ≤ ca < 90%) and seven substantially affected (ca < 50%) (Fig. S6B). Of these seven sensitive OCMs, 191, 109 and 246 were then further analysed in parallel with four resistant OCMs, namely 99, 105, 152 and 46-3 (Fig. 5A, B). First, we confirmed relative sensitivity, both in a colony formation assay in response to continuous PARGi exposure (Fig.  S6D, E), and in a time-lapse microscopy-based shortterm proliferation assay tracking GFP-tagged nuclei (Fig. 5C). Note that an N-methylated analogue of PARGi with minimal activity in vitro [28] was inactive against OCMs 191, 109 and 246 (Fig. S7), consistent with sensitivity due to on-target inhibition of PARG. Interestingly, there was no obvious correlation with cisplatin or paclitaxel sensitivity (Fig. 5D), and while OCMs 109 and 246 were resistant to the PARP inhibitors Olaparib and Niraparib (Fig. 5E, F), OCM.191 was PARPi-sensitive (Fig.  S6D, E). Indeed, the screen identified eight PARPi-sensitive OCMs, five of which were PARGi-resistant, while 191, 80-2, 59-3 were PARGi-sensitive (Fig. S6A-C). Thus, in summary, out of 32 patient-derived OCMs, we identified seven sensitive to the PARGi and eight sensitive to PARPi, including three sensitive to both.

PARG inhibition also induces replication catastrophe in sensitive OCMs
Having identified PARGi-sensitive OCMs, we asked whether they also showed the hallmarks of replication catastrophe upon PARG inhibition. As with the established cell lines, PARGi stabilised PAR chains in both PARGi-sensitive and -resistant OCMs in a PARPdependent manner (Fig. 6A, B), consistent with the ability of resistant cells to tolerate persistent PAR chains. Importantly, PARGi induced hallmarks of replication catastrophe in two of the three sensitive OCMs analysed (109 and 246), which was evidenced by significant increases in gH2AX, pKAP1, RPA foci, Chk1 phosphorylation and a pre-mitotic cell cycle block (Fig. 6C-H,  S8A, B). By contrast, the PARGi-resistant OCMs did not display these hallmarks. Thus, the two OCMs that are sensitive to PARGi, but PARPi-resistant, do indeed undergo replication catastrophe in a manner similar to the PARGi-sensitive established cell lines. Surprisingly, despite OCM.191 being sensitive to PARGi in both short and long-term assays, it did not display the features typical of replication catastrophe. A pre-mitotic block was evident in a small proportion of cells following PARGi treatment and, although most cells underwent mitosis, there were fewer total mitoses in PARGi-treated than control cells (Fig. S8A). This was partly owing to a longer interphase and, despite a similarly high rate of apoptotic events both with and without treatment, apoptosis tended to occur earlier in PARGi-treated cells. In addition, OCM.191 had an intermediate EC 50 , similar to OV56 in the cell line panel, which similarly showed moderate sensitivity to PARGi in short and long-term assays without features of significant replication stress with PARGi-treatment (Fig. 2D, S3).
Interestingly, OCM.191 is also PARPi-sensitive, and indeed it harbours a BRCA mutation and is possibly therefore HRD. Because HR defects may contribute to PARGi-sensitivity [22,24], one possibility therefore is that the defective HR may contribute to PARGi sensitivity in this case. Interestingly, OCM.246 was also derived from a patient with a germline BRCA2 mutation; however the patient received Olaparib maintenance monotherapy prior to biopsy sampling and the subsequent OCM harbours an intragenic BRCA2 reversion predicted to restore the open reading frame [51,52] (Supplementary Table 2), explaining ex vivo PARPi resistance. Consistently, OCM.246 appears to be HRP, evidenced by the accumulation of Rad51 foci in irradiated G2 cells (Fig. S8C). By contrast, in the same assay, OCM.109 appears to be HRD, consistent with the notion that HR status is not a predictor of PARGi sensitivity [26]. Thus, while the relationship between HR status, PARPi and PARGi sensitivity is complex, we nevertheless conclude that the PARGi sensitivity in patient-derived OCMs 109 and 246 is indeed due to the induction of replication catastrophe.

DNA replication gene expression in OCMs does not correlate with PARGi-sensitivity
Having demonstrated that PARGi-sensitivity was accompanied by replication catastrophe in two of the three sensitive OCMs examined, we asked whether the expression levels of DNA replication genes correlated with sensitivity, as observed in the established cell lines (Fig. 4B). The DNA replication, mitotic and apoptotic genes examined using NanoString in the established cell lines were evaluated in 29 OCMs for which we have RNAseq data [13,30,42]. As with our Nanostring data and the CCLE data, the expression of DNA replication and mitotic genes was highly correlated (Fig. S9A). In addition, there was a gradient in DNA replication expression across the OCM panel, with OCMs 86 and 106 showing the highest and lowest aggregated Z-scores respectively (Fig. 7A). However, ranking the OCMs by overall expression of the DNA replication genes did not differentiate PARGi-sensitive and -resistant OCMs (Fig. 7A, B). Similarly, ranking by mitotic and apoptotic genes also failed to differentiate PARGi sensitivity (Fig. S9B, C). This suggests that, in contrast to our earlier hypothesis [26], a DNA 'replication stress' gene expression signature may not be sufficient to serve as a predictive biomarker for PARG inhibitor sensitivity in a clinical cohort of HGSOC. Below we discuss these results in the context of developing a predictive biomarker for PARGi sensitivity.

Discussion
We set out to address two questions: (1) whether PARGisensitive ovarian cancer cells exhibit similar or different phenotypes in response to PARG inhibition, and (2) whether PARGi sensitivity correlates with the expression levels of DNA replication genes. Our analysis of 10 established ovarian cancer cell lines shows that indeed, the four PARGi-sensitive lines all displayed features of persistent replication stress upon exposure to PARGi. Moreover, interrogating three independent gene expression data sets identified a clear correlation between PARGi sensitivity and expression of DNA replication genes. We also observed a replication catastrophe phenotype in two PARGi-sensitive patient-derived ex vivo OCMs; however, in contrast to the established cell lines, they did not display lower DNA replication gene expression relative to PARGi-resistant OCMs. Below, we discuss these observations in terms of (a) the mechanisms responsible for PARGi sensitivity, and (b) efforts to develop biomarkers capable of predicting PARGi sensitivity.
To address the first question, we initially focused on a panel of 10 established ovarian cancer cell lines, four of which are PARGi sensitive and the remaining six, resistant. Upon PARG blockade, the four sensitive lines displayed features of replication stress and DNA damage, namely RPA and γH2AX foci, phosphorylation of Kap1 and Chk1, and asymmetric DNA replication forks, in turn leading to replication catastrophe illustrated by pannuclear γH2AX, pre-mitotic cell cycle arrest, reduced proliferation and suppressed outgrowth. By contrast, the resistant lines did not display these hallmarks, despite the stabilisation of PAR chains. Of the three PARGi-sensitive OCMs we analysed in detail, two also exhibited hallmarks of replication catastrophe, extending our observations from established cell lines to patient-derived ovarian cancer cells that have not undergone extensive ex vivo proliferation [13]. Together, this recurrent replication catastrophe phenotype supports our prior hypothesis that PARGi-sensitivity arises due to a pre-existing DNA replication vulnerability that prevents toleration of stabilised PAR chains. The nature of this vulnerability remains unclear but could reflect compromised replisome function, a notion supported by reduced DNA replication gene transcripts in sensitive lines. However, whether these genes are down-regulated due to active suppression of transcription is not clear. Alternatively, resistant cell lines may have actively up-regulated DNA replication genes, possibly as an adaptive response to oncogene-induced replication stress [53,54]. In budding yeast, various adaptive mechanisms can increase fitness in response to replication stress [55], raising the possibility that differential PARGi sensitivity reflects different responses to oncogenic replication stress. Thus, while adaption via up-regulating DNA replication genes could engender intrinsic PARGi resistance, adaptation via other mechanisms may leave cells vulnerable to PARG blockade. Exploring the evolution of acquired PARGi resistance may shed light on this issue.
While our data strongly points to a DNA replication vulnerability, other mechanisms can also render cells dependent on PARG activity. By sequestering NAD+ in stabilised PAR chains, PARG inhibition was recently shown to cause metabolic catastrophe in glioma cell lines harbouring IDH1 mutations [56]. This study employed the same PARG inhibitor but with sensitivity observed at a five-fold higher drug concentration and only when used in combination with temozolomide, indicating that in this case, sensitivity is due to a 'triple whammy' effect requiring IDH1 mutation, DNA damage and PAR stabilisation. Interestingly, in our study, partial PARGi sensitivity in OV56 was not accompanied by overt replication stress. Similarly, PARGi-sensitive OCM.191 did not show signs of replication stress with treatment. Whether this intermediate sensitivity is accounted for by NAD+ sequestration and metabolic catastrophe remains to be seen. Moreover, whether PARGi-induced metabolic catastrophe is distinct from or related to the DNA replication vulnerability mechanism is also unclear. Indeed -because, firstly NAD+ metabolism is required to produce NADP+, and subsequently nucleic acid precursors via the pentose-phosphate pathway [57,58], and secondly nucleotide depletion increases replication stress -PARGi sensitivity could reflect a complex interplay between DNA replication and NAD+ vulnerabilities. For example, at low doses, PARGi may deplete NAD+ sufficiently to perturb nucleotide synthesis resulting in replication stress that can only be buffered by cells that have adapted to oncogene-induced replication stress. By contrast, when combined with acute DNA damage at higher doses, PARGi may deplete NAD+ to the point that it perturbs NAD+ pathways more broadly, leading to metabolic catastrophe. An interesting corollary therefore is that PARGi-sensitising DNA replication vulnerabilities could arise due to upstream defects in supplying nucleotides to the replisome, which are exacerbated by PARG blockade. Exploring the interplay between NAD+ metabolism, nucleotide synthesis and DNA replication in the context of PAR dynamics will be important next steps.
Our second question regarding a potential correlation between PARGi sensitivity and expression levels of DNA replication genes was motivated by the quest to develop predictive biomarkers for PARGi sensitivity to enable patient stratification for clinical trials. Interrogating 25 DNA replication genes in two independent RNAseq datasets and our NanoString analysis shows a clear correlation; the four PARGi-sensitive ovarian cancer cell lines show reduced expression levels versus the resistant lines. If this correlation extended to patient-derived OCMs, then one could envision a 'replication stress' gene expression signature providing the basis of a biomarker [26]. However, we did not observe a similar effect in PARGisensitive OCMs. There are various possible explanations for the non-concordance between established cell lines and OCMs. One possibility is sample size; while the colony formation screen identified seven OCMs as potentially sensitive, only OCMs 109 and 246 were particularly sensitive with nanomolar EC 50 values. As our living biobank grows, it will be important therefore to identify additional PARGi-sensitive OCMs to further test whether DNA replication gene expression correlates with sensitivity. A second possibility is that analysing established cell lines is complicated by different ovarian cancer subtypes. While PARGi-sensitive Kuramochi is HGSOCderived, the three other sensitive cell lines (OVMANA, OVISE and RMG1) represent the clear cell subtype [30,45]. Thus, the correlation may reflect a clear cell phenomenon not shared more broadly with HGSOC. Interestingly, OCMs 80-2 and 87 may be clear-cell-derived [30], and while OCM.80-2 is PARGi sensitive and has reduced DNA replication transcripts, OCM.87 is more resistant and does not, reinforcing the need to analyse PARGi sensitivity in a broad range of OCMs.
Another complication extending observations from established cell lines to OCMs is chromosome instability (CIN). Although ovarian cancer cell lines exhibit CIN, it is markedly reduced compared with the mitotic chaos observed in primary cultures, most probably because extended ex vivo propagation selects fitter, relatively stable clones [13,59]. Constant chromosome reshuffling likely generates transcript heterogeneity, in turn influencing drug sensitivity and the stability of gene expression signatures [60][61][62]. Interestingly, the two PARGi-sensitive OCMs that displayed PARGi-induced replication stress (109 and 246) have karyotypes dominated by relatively stable monosomies and disomies, compared with the highly variable chromosome gains and focal amplifications typical of resistant OCMs (Fig. S5F). Exploring the relationship between karyotypic features, CIN and PARGi sensitivity will also be facilitated by a broad range of well-characterised OCMs.
And finally, the correlation between replisome gene expression and PARGi sensitivity is partly a circular argument. Our initial focus on replisome genes was motivated by (a) the replication catastrophe phenotype, and (b) the emergence of replisome genes in a PARGi-sensitiser RNAi screen [26]. In turn, interrogating replisome genes led us to test sensitivity of OVMANA, OVISE and RMG1, which we now confirm have relatively low expression levels of DNA replication genes. This highlights the need to take fresh, unbiased approaches on larger collections of OCMs with defined drug sensitivity. Indeed, it is striking that the expression of S-phase and mitotic genes correlates strongly, not only in the panel of 10 established cell lines but generally across the CCLE dataset and our cohort of OCMs. This raises the possibility that the differential gene expression between PARGi-sensitive and -resistant cell lines is mediated by a common upstream regulatory network, e.g. the E2F network, which drives both S-phase and mitotic gene expression programs [63]. Indeed, E2F network genes, along with DNA replication and mitotic spindle genes, are up-regulated following introduction of TP53 and BRCA1 mutations in a cell line model of ovarian cancer development, corresponding with the emergence of CIN [64]. Alternatively, correlated expression of S-phase and mitotic genes may independently arise in response to the same selective pressure, namely oncogene-induced replication stress. The stronger correlation observed for our cell line panel, than in cancer cell lines generally, in a disease characterised by high levels of CNV, implies that a common evolutionary pressure is a more likely explanation. Interestingly, in the yeast model alluded to above, in addition to DNA replication processes, sister chromatid cohesion networks were also repeatedly altered by exposure to replication stress [55]. PARGi-sensitivity may therefore reflect a broader adaptation of cell cycle processes that arise in response to oncogene-induced deregulation of cell cycle controls.
Synthesising these issues, we conclude that the relationship between the expression levels of replisome genes and PARGi sensitivity is more complex than initially proposed [26]. Therefore, developing robust predictive biomarkers with potential for clinical utility will require more mechanistic insight, driven by both hypothesisled and unbiased approaches. Nevertheless, a key outcome of this study is the identification of two OCMs that are particularly sensitive to single-agent PARG inhibition. While we previously described OCMs sensitive to PARGi in combination with a CHK1 inhibitor [26], these latter observations indicate that PARG inhibitors may have efficacy as monotherapies. This is encouraging because there is a need for new therapeutic agents, with appropriate predictive biomarkers, and in particular for the large cohort of women with HRP HGSOC who are unlikely to benefit from PARP inhibitors. A key question therefore is whether PARG inhibitors will offer distinct therapeutic opportunities to PARP inhibitors. Interestingly, our screen identified a similar proportion of OCMs with PARGi and PARPi sensitivity, and of seven PARGisensitive OCMs, four were PARPi-resistant, consistent with previous reports that PARGi and PARPi sensitivity are mostly non-overlapping [26,28]. Indeed, of the six PARGi-resistant established cell lines used in this study, four are sensitive to Olaparib, and of four PARGisensitive cell lines, only two are sensitive to Olaparib.
Non-overlapping sensitivity to PARPi and PARGi may seem counter-intuitive, as PARP1/2 and PARG work in concert to repair DNA damage; one might expect that both PARPi and PARGi would be toxic towards tumour cells with defects in DNA damage repair [22]. However, while some studies suggest synthetic lethality between BRCA1/2 mutations and PARG inhibition [22,24] others do not [26,65]

Conclusions
Here we show that in a panel of established cell lines, PARGi sensitivity is accompanied by hallmarks of replication catastrophe, and correlates with lower expression of DNA replication genes. We also identify several patient-derived OCMs that are sensitive to PARGi monotherapy, again via a DNA replication catastrophe mechanism. However, DNA replication gene expression did not correlate with sensitivity of the OCMs meaning a DNA 'replication stress' gene expression signature is unlikely to be a sufficient predictive biomarker for PARG inhibitor sensitivity in a clinical cohort of HGSOC. These results highlight the complexity of developing a predictive biomarker for PARGi sensitivity. Thus, while further research is required to delineate mechanisms of PARGi sensitivity and to develop predictive biomarkers, this study nonetheless reinforces the potential of PARG as a new therapeutic target for women with HGSOC, including those who develop platinum-resistant disease who currently have an overall survival of only ~ 12 months [68]. Significantly, OCMs 109 and 246, the two most highly PARGi-sensitive, were biopsied from women with platinum-resistant disease. Moreover, although patient 246 was previously treated with Olaparib maintenance monotherapy, OCM.246 has a BRCA2-reversion, is HRP and PARPi-resistant. Thus, this study indicates that PARG inhibitors may represent a future alternative treatment for patients with otherwise limited therapeutic options, such as those with disease that is platinum and/ or PARPi resistant, and informs the design of early clinical studies of PARG inhibitors.  [39]. Ten cell line panel indicated by black diamonds, note OV56 CCLE expression data does not reflect NanoString expression data (see text). Pearson R 2 values indicated below the graph, all comparisons, p < 0.0001, total number cell lines 747. ns = not significant, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. values. Data are mean ± SEM from 3 biological replicates. PRISM could not accurately calculate EC 50 for PARGi-resistant cells, therefore for resistant OCMs, the EC 50 was approximated as 50 μM (half the maximal concentration tested) and described in the table as not calculable (NC).