ADP-Ribosylation Levels and Patterns Correlate with Gene Expression and Clinical Outcomes in Ovarian Cancers

Ovarian cancer is the 10th most common cancer among women in the United States (incidence of ∼22,000 cases of ovarian cancer diagnosed per year), but is the 5th leading cause of cancer mortality in women (∼14,000 deaths per year; refs. 1, 2). Ovarian cancer is the most deadly of all gynecologic cancers, with less than 50% of patients surviving 5 years (3). Because early-stage ovarian cancer is difficult to detect, most patients have advanced disease (stage III–IV) at diagnosis (4). Despite advances in radical surgery and chemotherapy for the treatment of advanced ovarian cancers, up to 85% eventually relapse and response to subsequent cytotoxic therapies is short-lived (5).

The first molecular aberrations to be targeted therapeutically in ovarian cancers were germline mutations in genes encoding components of homologous recombination-mediated DNA repair (HRR) pathways (e.g., BRCA1, BRCA2). In 2005, two landmark studies demonstrated the effectiveness of PARP inhibitors in BRCA1/2-mutated cancers in preclinical models (6, 7). These studies demonstrated efficacy through a mechanism involving synthetic lethality between the PARP inhibitors, which block PARP-mediated DNA repair, and the HRR pathways controlled by BRCA1 and BRCA2 (6, 7). In 2014, the U.S. Federal Drug Administration (FDA) approved the use of the PARP inhibitor Olaparib (Lynparza) as a 4th line agent for recurrent ovarian cancers harboring deleterious germline mutations in BRCA1 or BRCA2 (8–10). The FDA has since expanded the use of PARP inhibitors, including Rucaparib (Rubraca) and Niraparib (Zejula), to germline BRCA1/2-mutated patients with recurrent epithelial ovarian, fallopian tube, or primary peritoneal cancers after 2 lines of platinum-based chemotherapy (11, 12).

Evidence of potential efficacy of PARP inhibitors in nongermline BRCA1/2-mutated ovarian cancers has led to FDA approval of PARP inhibitors for use in certain clinical scenarios regardless of BRCA1/2 status (13). Given the effectiveness of PARP inhibitors in the HRR pathway, there has been a considerable effort made to develop biomarkers to identify an HRR dysfunction phenotype that could predict PARP inhibitor sensitivity. In addition, given (i) emerging roles of nuclear PARP proteins in molecular pathways beyond DNA repair (e.g., transcription, chromatin regulation, RNA processing; refs. 14, 15) and (ii) potential therapeutic effects of PARP inhibitors in cancers without BRCA1/2 mutations or observable HRR defects (11, 13, 16), biomarkers for PARP inhibitor responsiveness in these cases are needed as well.

Most of our understanding of the molecular and biological functions of PARPs and the posttranslation modification of proteins that they mediate [i.e., ADP-ribosylation (ADPRylation)] has come from studies with nuclear PARPs, in particular PARP-1, the most abundant and ubiquitous member of the PARP family (14, 15). PARP-1 is a ubiquitously expressed nuclear enzyme that has conserved functional domains for both nonsequence-specific DNA binding and nicotinamide adenine dinucleotide (NAD⁺)-dependent catalytic activity (14, 15). Upon activation, PARP-1 uses NAD⁺ to catalyze the addition of PAR polymers on itself (i.e., automodification) and other substrate proteins (17, 18). Other PARP family members catalyze the addition of mono(ADP-ribose) (MAR) monomers (14, 18). The levels of protein-linked PAR and MAR are a direct indication of the levels of PARP activity in the cell, as well as the biology of the cell. While few studies have investigated endogenous ADPRylation as a potential biomarker for responses to PARP inhibitor treatment, some progress has been made in this regard, with correlations between the levels of PARylation and responses to PARP inhibitors observed (19, 20).

Currently, the most widely used tool for the detection of PARylated proteins is a monoclonal antibody (10H mAb), which detects polymers of ADP-ribose greater than nearly 10 to 15 units in length (21). The 10H mAb, however, does not detect other products of PARP activity, such as oligo(ADP-ribose) (OAR) and MAR, which are likely to have important biological roles. Our lab has recently developed a set of recombinant antibody-like ADP-ribose detection reagents that can recognize all forms of ADP-ribose, including: (i) pan-ADP-ribose (MAR/OAR/PAR) and (ii) OAR/PAR (22). Various immunological techniques have been developed for detecting PARP-1 proteins and activity levels in cultured cells or clinical tumor samples (23–30). In studies described herein, we used our recombinant antibody-like ADP-ribose detection reagents to characterize the levels and patterns of ADPRylation in a cohort of ovarian cancer tumor tissue samples and relate the results to molecular, cellular, and clinical outcomes. Our results indicate that ADPRylation can be a useful predictive biomarker in ovarian cancers.

We obtained tissue samples intraoperatively from surgical tumor specimens of patients with high-grade serous ovarian cancer at the time of the planned cytoreductive surgery under a tissue repository protocol from 2009 to 2014. Samples were immediately flash frozen in liquid nitrogen, and stored at −80°C until use. Samples from 34 different patients were used in this study. Patient data were obtained from electronic medical records and included: age, race, stage, histology, laboratory values, debulking status, germline BRCA mutation status, chemotherapy regimens, platinum status, and survival outcomes. These studies were conducted in accordance with recognized U.S. and international ethical guidelines, and informed consent was obtained from the patients. This study was performed with approval by the University of Texas Southwestern Medical Center Institutional Review Board (IRB No. #STU-012015-019).

Primary antibodies used for Western blotting were as follows: (i) recombinant antibody-like PAR binding reagent (WWE-Fc; anti-PAR) and pan ADP-ribose binding reagent (Macro-Fc; anti-panADPR), which were expressed in bacteria and purified as previously described (ref. 22; now available from EMD Millipore; catalog nos. MABE1031 and MABE1016, respectively) and (ii) rabbit anti-PARP-1 polyclonal antiserum (generated in the Kraus lab; ref. 31), used at 1:1,000 dilution. Primary antibodies against SNRP70 (Abcam ab83306; 1:4,000 dilution) and TFIID (SI-1; Santa Cruz Biotechnology, sc-273; 1:500 dilution) were used for detection of loading controls. The secondary antibody used for detection was a goat anti-rabbit HRP-conjugated antibody (Thermo Fisher Scientific, 31346; 1:2,500 dilution).

Frozen tumor samples were homogenized in 1× isotonic lysis buffer (50 mmol/L Tris-HCl pH 7.5, 10 mmol/L MgCl₂, 15 mmol/L CaCl₂, 1.5 mmol/L sucrose) supplemented with 1 mmol/L dithiothreitol (DTT; Sigma), 1× protease inhibitor cocktail (Sigma), 250 U/μL Universal Nuclease (Thermo Fisher Scientific), 20 μmol/L PJ34 (Sigma), and 500 nmol/L adenosine 5′-diphosphate (hydroxymethyl)pyrrolidinediol (ADP-HPD; Millipore) using a Bullet Blender (Next Advance). The homogenate was subjected to centrifugation for 20 minutes at full speed in a microcentrifuge at 4°C, and the supernatant containing the cytoplasmic extract was collected, flash-frozen, and stored at −80°C for later use. Nuclear extracts were prepared from the pelleted nuclei by adding extraction buffer [20 mmol/L HEPES pH 7.9, 1.5 mmol/L MgCl₂, 0.42 mmol/L NaCl, 0.2 mmol/L EDTA, 25% (v/v) Glycerol] supplemented as described above, following a protocol provided with the CelLytic NuCLEAR Extraction Kit (Sigma). The protein concentrations of the nuclear extracts were determined using a Bradford Assay Kit (Thermo Fisher Scientific) and the remaining nuclear extracts were aliquoted, flash-frozen, and stored at −80°C for later use.

Ten micrograms of nuclear extract protein from each sample was prepared for SDS-PAGE by adding 4× SDS sample buffer to a final concentration of 1×, followed by incubation for 5 minutes at 100°C. The samples were resolved on 10% PAGE-SDS gels for ADPR and 6% PAGE-SDS gels for PARP-1. The protein gels were transferred to a nylon-backed nitrocellulose membrane. The membranes were blocked for 1 hour at room temperature in tris-buffered saline with 0.1% Tween 20 (TBST) containing 5% nonfat dry milk. Primary antibodies were diluted in 1% nonfat dry milk and incubated for 1 hour at room temperature. After extensive washing in TBST, the membranes were incubated with an HRP-conjugated secondary antibody diluted in 1% nonfat dry milk for 1 hour at room temperature. Signals were revealed using a luminol-based enhanced chemiluminescence substrate (SuperSignal West Pico; Thermo Fisher Scientific) and a ChemiDoc imaging system (Bio-Rad).

ADPR signal intensities from Western blots performed using the Macro-Fc and WWE-Fc reagents were quantified utilizing ImageJ 1.50i software. We standardized the blots based on total protein loaded per lane, which was verified after Western transfer by staining the membrane with amido black. We controlled for background intensity and determined the relative signal intensity for each Western blot lane (single patient or cell line). We normalized between blots and replicates by equalizing the PAR signals from Sample 1, which were consistently among the highest that we obtained with both the Macro-Fc and WWE-Fc reagents. The ratios of the Macro-Fc and WWE-Fc signals were determined by dividing the normalized signals for Macro-Fc by the normalized signals from the WWE-Fc reagent. Similar methods were applied to the PARP-1 signal band of each patient or cell line to determine the relative signal intensity.

Total RNA was isolated from each ovarian cancer tissue sample using an RNeasy Plus Kit (Qiagen), with QIAshredder spin columns for homogenization and on-column DNaseI digestion (79254; Qiagen). The total RNA was analyzed for integrity using an Agilent 2100 Bioanalyzer. The RNA collected was then processed for whole-genome polyadenylated RNA sequencing (polyA+ RNA-seq).

The total RNA samples were enriched for polyA+ RNA using Dynabead Oligo(dT)25 (Invitrogen) and strand-specific RNA-seq libraries were prepared from the polyA+ RNA as described previously (32). The RNA-seq libraries were sequenced using an Illumina HiSeq 2500 (single-end, 50 nt; SE50; one replicate). The RNA-seq data files are available from the NCBI's Gene Expression Omnibus (GEO) portal using accession number GSE136155.

We used a computational pipeline to analyze the RNA-seq data, which includes the following steps:

• (1) RNA-seq read mapping. Single-end RNA-seq reads were aligned to the human reference genome (NCBI 37, hg19) using the spliced read aligner TopHat version 2.0.13 (33) with default parameters.

• (2) Transcriptome assembly. The transcriptome for each sample was then assembled using Cufflinks v.2.2.1 (34). We ran Cufflinks using the transcript abundance calculation mode to estimate the read coverage of each transcript.

• (3) Differential expression. The assembled transcripts from each individual library were merged using Cuffmerge to generate a master transcriptome. The merged transcripts were then used to determine the FPKM for all transcripts in the patient samples using Cuffdiff.

• (4) Filtering transcripts based on a transcript abundance threshold. We obtained transcript abundance, in terms of FPKM, for all transcripts from Cufflinks. A FPKM threshold of 10 was applied to all samples to limit our focus on those transcripts that are reasonably expressed in ovarian cancer tissue samples.

The uniquely mapped reads were converted into bigWig files for visualization in a UCSC genome browser. The writeWiggle function from the groHMM (35) package was used to convert the counts obtained using GRanges to the wiggle file format. The wiggle files were converted to bigwig using wigToBigWig function of the UCSC Genome Browser.

The FPKM values obtained from the differential expression analysis of the PARP1 mRNA from RNA-seq were compared with the relative signals from the PARP-1 Western blots (quantified as described above) and used to generate a scatter plot. A similar analysis was performed for the relative signals from the Macro-Fc Western blots versus the relative signals from the PARP-1 Western blots.

The deviation of expression for each gene in each patient from the mean expression of the gene across all patients was calculated and used to make the heat map. Only those genes with significant deviations were included. The FPKM of the transcripts obtained were subjected to hierarchical clustering using Cluster 3.0 (36) with average linkage to cluster the samples based on the similarity of the expression pattern.

Statistical analyses were performed using GraphPad Prism 6. Quantification of Macro-Fc, WWE-Fc, and PARP-1 signal intensities, as well as RNA-seq FPKM, for patients within each of the 4 groups were analyzed using a 1-way ANOVA Tukey–Kramer multiple comparison test. Pearson's correlations were computed to compare Macro-Fc PAR signal intensity to PARP-1 protein expression level, and for PARP-1 protein expression level to PARP1 mRNA RNA-seq FPKM.

Progression-free survival (PFS) was defined as the time from the date of diagnosis to the date of first recurrence. Overall survival (OS) was defined as the time from the date of diagnosis to the date of death from disease or last follow-up. Kaplan–Meier analyses (37, 38) were used to determine survival outcomes and differences were analyzed with the log rank (Mantel–Cox) test. A P-value < 0.05 was considered significant.

Patient characteristics, including age and pretreatment CA-125 levels (units/mL) were compared using 1-way ANOVA Tukey–Kramer multiple comparison test. The number of patients in each group were compared using a Student t test. Stage at diagnosis (III vs. IV), type of frontline chemotherapy (adjuvant, neoadjuvant, none), surgical residual disease (no gross, optimal, suboptimal), platinum status (sensitive, resistant, refractory), and germline BRCA mutation status [negative, BRCA1 deleterious, BRCA2 deleterious, BRCA variant of unknown significance (BRCA VUS), and unknown] were compared using the Chi-square method.

OVCAR-3 and OVCAR-4 cells were purchased from the ATCC and were cultured as recommended. HCC5012, HCC5022, HCC5023, HCC5024, HCC5044, and HCC5048 cells were described previously (39) and were obtained directly from the source of the cells (Dr. Adi Gazdar). They were cultured in RPMI1640 medium supplemented with 10% FBS and 2 mmol/L l-glutamine. All cells were used at early passage and were replenished after no more than 15 passages. All cells were verified as mycoplasma-free every 3 months using a PCR-Based Screening Kit.

Whole cell lysates (WCL) and nuclear extracts were prepared from each ovarian cancer cell line. WCLs were prepared using a high salt extraction buffer [50 mmol/L Tris, pH 7.5, 500 mmol/L NaCl, 1 mmol/L EDTA, 1% NP-40, 10% glycerol, 1 mmol/L β-mercaptoethanol, 1× protease inhibitor cocktail; Roche] containing 10 μmol/L of PJ34 (Sigma) and 500 nmol/L of ADP-HPD (Millipore). After cell lysis, extraction on ice for 30 minutes, and clarification by centrifugation for 20 minutes at full speed in a microcentrifuge at 4°C, the extracts were collected and flash-frozen in liquid N₂. Aliquots were used for determination of protein concentrations using a Bradford assay (Bio-Rad).

Nuclear extracts were prepared using isotonic lysis buffer (10 mmol/L Tris pH 7.5, 2 mmol/L MgCl₂, 3 mmol/L CaCl₂, and 300 mmol/L sucrose, 1 mmol/L DTT, 1× protease inhibitor cocktail; Roche) containing 10 μmol/L of PJ34 and 500 nmol/L of ADP-HPD. After swelling the cells on ice in isotonic lysis buffer for 15 minutes, cell lysis was induced using 0.6% NP-40. The nuclei were collected by centrifugation on low speed in a microfuge at 4°C and then incubated with extraction buffer (20 mmol/L HEPES, pH 7.9, 1.5 mmol/L MgCl₂, 0.42 mmol/L NaCl, 0.2 mmol/L EDTA, 25% glycerol, 1 mmol/L DTT, 1× protease inhibitor cocktail; Roche) containing 10 μmol/L of PJ34 and 500 nmol/L of ADP-HPD at 4°C for 30 minutes with constant shaking. The nuclear extracts were clarified by centrifugation for 20 minutes at full speed in a microcentrifuge at 4°C, collected and flash-frozen in liquid N₂. Aliquots were used for determination of protein concentrations using a Bradford assay (Bio-Rad).

For detection of ADPRylation in WCLs and nuclear extracts, 20 μg of lysate from each cell line was used for 8% PAGE-SDS gel electrophoresis. After transfer to a nylon-backed nitrocellulose membrane, the membranes were blotted using the Macro-Fc and WWE-Fc reagents as described above.

The ovarian cancer cell lines noted above were seeded in 96-well polystyrene clear bottom plates (Costar; Catalog No. 3610) and treated with Olaparib for 48 hours, as indicated. Olaparib (MedChem Express, HC-10162) was prepared in DMSO and 1:5 serial dilutions of the drug were prepared. Cell survival was quantified using CellTiter-Glo Luminescent Cell Viability Assay (Promega). Luminescence was measured using Tecan Spark multimode microplate reader (Tecan Trading AG). All experiments were performed with 3 or more biological replicates to ensure reproducibility.

Given the importance of PARP-1 in the biology of ovarian cancer and as a therapeutic target (40–42), we hypothesized that the product of PARP activity (i.e., covalent ADPRylation of substrate proteins) might be a good indicator of PARP-1 function in ovarian cancers. We collected tumor tissue from 34 patients with high-grade serous ovarian cancers (the clinical characteristics of the patients and their cancers are presented in Table 1 and Supplementary Table S1). We used these samples to determine selected molecular phenotypes of the cancers and associate them with patient clinical outcomes (Supplementary Fig. S1).

To determine the relative levels and patterns of protein-linked ADPRylation in these samples, we performed Western blotting of nuclear extracts using 2 recombinant antibody-like ADPR detection reagents previously developed in our lab: (i) Macro-Fc, which detects MAR, OAR, and PAR (pan-ADPR), and (ii) WWE-Fc, which detects, OAR and PAR (ref. 22; Fig. 1A). We also monitored PARP-1 protein levels by Western blotting and quantified the signal intensities in arbitrary units for ADPR (from Macro-Fc and WWE-Fc,) and PARP-1 using ImageJ software. We then determined the levels of the Macro-Fc and WWE-Fc signals, as well as their ratios by dividing the normalized signals for each reagent.

To categorize the samples into groups based on the levels and patterns of ADPRylation, we attempted to integrate various parameters from the Western blots, including: (i) WWE-Fc signal, (ii) macro-Fc signal, (iii) macro-Fc/WWE-Fc ratio, (iv) PARP-1 signal, and (v) banding pattern. Initially, we tested a variety of unbiased clustering algorithms using one or more of the quantitative parameters listed above. These analyses yielded a disparate set of results that varied with respect to both the number of clusters and how the samples were distributed within the clusters. Adding some of the clinical parameters from the patients to the clustering analyses did not improve the consistency.

Thus, we developed a rubric based on cut-offs applied to the WWE-Fc signals and the macro-Fc/WWE-Fc ratio (Supplementary Fig. S2A), which was also guided, in part, by (i) the banding patterns of the ADP-ribose signals, (ii) the PARP-1 signal intensity, and (iii) the most consistent aspects of the unbiased clustering described above (e.g., separation of the samples that comprise groups 1 and 2). This led to the 4 groups of samples shown in Fig. 1A. We validated aspects of this approach by using hierarchical clustering with a limited set of the quantitative parameters derived from the Western blots (i.e., WWE-Fc signal, macro-Fc signal, macro-Fc/WWE-Fc ratio). This analysis separated group 1 from the other groups (Supplementary Fig. S2B), as well as group 3 from group 2 at more granular levels of the tree (Supplementary Fig. S2C). Unlike the rubric, this approach did not separate group 4 from group 2 (Supplementary Fig. S2C). For subsequent analyses, we kept the groups 3 and 4 samples distinct from groups 1 and 2, rather than combining them, because of their distinctive patterns, signal intensities, and ratios of macro-Fc/WWE-Fc signals (Fig. 1A). Although a number of patient samples had a known deleterious germline mutation in BRCA1 or BRCA2 (denoted with an asterisk), these samples were not enriched in any group (Fig. 1A; Supplementary Table S1).

We analyzed the quantitative differences in the Western blot signal intensities across the 4 groups using one-way ANOVA Tukey–Kramer multiple comparison tests (Fig. 1B–E). The ADPR signals from the Macro-Fc reagent were most intense for Group 1, but did not differ significantly among Groups 2, 3, or 4 (P < 0.01; Fig. 1B). The ADPR signals from the WWE-Fc reagent showed greater variability among the groups, with the greatest intensity for group 1 (P < 0.001; Fig. 1C). The ratio of signals from the macro-Fc and WWE-Fc reagents was significantly different for group 3 (P < 0.0001; mean ratio 4.26), as expected, but were not significantly different for group 1 (mean ratio 0.999), group 2 (mean ratio 1.07), and group 3 (mean ratio = 0.46; P < 0.0001; Fig. 1D). PARP-1 expression levels were significantly greater in group 1 than groups 2, 3, or 4 (P < 0.0002), but the expression was not significantly different among groups 2, 3, or 4 (P > 0.25; Fig. 1E).

Finally, we assessed the potential contribution of PARP-1, the most abundant and active nuclear PARP, to the ADPR signal observed. To do so, we plotted ADPR signal intensity from the macro-Fc reagent versus PARP-1 expression level for all tumor samples (n = 34). Using Pearson correlation analysis, we observed a significant positive correlation between ADPR and PARP-1 levels (Fig. 2A, Pearson r = 0.530, P = 0.009). These results suggest that higher levels of ADPRylation are driven by higher PARP-1 protein levels in ovarian cancers. Together, the relative quantification and comparison of ADPR and PARP-1 levels illustrates the molecular differences among the 4 groups.

To better understand the biological meaning of ADPRylation in ovarian cancers, we sought to relate ADPR levels to gene expression, another frequently used indicator of cancer cell biology. We subjected a subset of patient samples from each group (n = 22 total) to RNA sequencing (see Supplementary Table S1 for the samples selected). For our initial analysis using the RNA-seq data, we determined the relationship between PARP1 mRNA and PARP-1 protein levels. We plotted the mapped reads for PARP1 mRNA from RNA-seq (fragments per kilobase of transcript per million, FPKM) versus the PARP-1 protein levels from Western blotting (arbitrary units; Fig. 2B). Using Pearson correlation, we observed a significant positive correlation between PARP1 mRNA and PARP-1 protein levels (Pearson r = 0.603, P < 0.001). Thus, as expected, higher PARP1 RNA levels correlated positively with higher PARP-1 protein levels.

Next, we assessed the patterns of gene expression across the ovarian cancer samples using the RNA-seq data. We calculated the deviation of expression for each gene in each ovarian cancer sample from the mean expression of the gene across all samples. Only those genes with significant deviations from the average (at least 1 SD) were included, yielding a total of 6,730 significant differentially regulated genes. The FPKM of the transcripts obtained were subjected to hierarchical clustering based on the similarity of the expression patterns. The data were then presented in the heatmap shown in Fig. 3A. We also averaged the FPKMs for each gene in the 4 previously derived ADPRylation groups and presented them in the heatmap shown in Fig. 3B. Finally, we quantified the expression from the upper (U) and lower (L) one-third of genes represented by the aforementioned heatmap and expressed the results in the boxplots shown in Fig. 3C. We observed significant differences across the groups (Wilcoxon rank sum test, P value <2.2 × 10⁻¹⁶). Together, these results demonstrate that the gene expression patterns differ dramatically across the 4 ADPRylation groups, suggesting that ADPR levels track with the biology of the cancer types.

We used unbiased clustering of the RNA-seq data to determine if we could observe sample groupings similar to those derived based on the levels and patterns of ADPRylation. We found that k-means clustering separated groups 1 and 2, and to a lesser extent group 3, similar to the rubric and hierarchical clustering used for the Western blot data (Supplementary Fig. S3A). This approach, however, did not separate group 4 from group 2 (Supplementary Fig. S3B), similar to what we observed with the hierarchical clustering used for the Western blot data (Supplementary Fig. S2C). Thus, multiple aspects of our clustering analyses indicate that the ADPRylation and the gene expression results are mutually supportive when it comes to categorizing the samples.

Finally, we defined gene ontologies from the upper (U) one-third and lower (L) one-third of significantly expressed genes (Fig. 3D). The genes most highly expressed in group 1 are involved in transcriptional regulation and cellular signaling, whereas the genes most highly expressed in Groups 2 to 4 are involved in cell adhesion, extracellular organization, and inflammatory responses. The gene ontologies for the latter are reminiscent of aggressive cancers, which rearrange their extracellular matrices to alter their ability to adhere and invade surrounding tissues. Collectively, the gene expression analyses described here show that molecular readouts, such as gene expression and ADPRylation levels, can reveal information about cancer cell biology.

Next, we sought to determine if the levels and patterns of ADPRylation in the 4 groups defined above correlate with clinical outcomes. We plotted Kaplan–Meier recurrence-free (RFS) and overall (OS) survival curves for each of the 4 groups (Fig. 4A and B). After a median follow-up of 23.5 months, group 1 exhibited significantly better RFS compared with group 4 (median RFS of 22 months vs. 11 months, respectively P = 0.009; Fig. 4A and C). Although the median OS was not reached (NR) for groups 1, 2, and 3; the OS was significantly better for group 1 compared with group 4 (median OS, not reached versus 37 months, respectively, P = 0.04; Fig. 4B and C). Together, these results indicate that the levels and patterns of ADPRylation are indicative of clinical outcomes in ovarian cancers in response to standard-of-care platinum-based therapy.

We note that for group 4, the debulking surgery was optimal in only 25% (1/4) of the patients, compared with 67% (6/9) in group 1, 47% (7/15) in group 2, and 50% (3/6) in group 3 (Supplementary Table S1). Because the status of the debulking surgery is an important prognostic factor in high-grade serous ovarian cancer, we cannot rule out its contribution to the differences seen in RFS with group 4. But, we also we do not have more detailed surgical information to explain additional factors that may have led to suboptimal debulking, which may be complicating covariates.

Given the growing importance of PARPi in the treatment of advanced ovarian cancers, as well as the correlations between the levels of ADPRylation and clinical outcomes, as shown in Fig. 4, we were interested in determining whether the levels of ADPRylation might also relate to the response of ovarian cancers to PARPi. To explore this possibility, we used 8 patient-derived, high-grade serous ovarian cell lines [HCC lines 5012, 5022, 5023, 5024, 5044, 5048 (39); OVCAR-3, OVCAR-4]. These cells lines exhibited a range of levels of ADPRylation (Supplementary Fig. S4A) and ratios of the signals from macro-Fc and WWE-FC (Supplementary Fig. S4B). Only 2 of the 8 cell lines harbor potential deleterious mutations in BRCA1/2 (Supplementary Fig. S4A, yellow box). To evaluate PARPi sensitivity, we treated the cell lines with Olaparib, an FDA-approved PARPi, at various concentrations for 48 hours (Fig. 5A). We then determined cell survival for all 8 cell lines using a luminescent cell viability assay (Supplementary Fig. S5).

OVCAR-3, which exhibited the highest levels of ADPRylation (Fig. 5B; Supplementary Fig. S4A), was the most sensitive to Olaparib (EC₅₀ ∼1 μmol/L; Fig. 5C). In contrast, HCC5048, which exhibited the lowest levels of ADPRylation (Fig. 5B; Supplementary Fig. S4A), was largely insensitive to Olaparib (Fig. 5C). The relationship was not limited to these 2 cell lines; across the 8 cell lines, sensitivity to Olaparib was strongly and directly correlated with the levels of ADPRylation (in other words, the EC₅₀s for Olaparib were strongly and inversely and correlated with the levels of ADPRylation; Pearson R = −0.86, P-value = 0.0068; Fig. 5B and D; Supplementary Figs. S4–S6). As expected, treatment with Olaparib inhibited ADPRylation in the cells (Supplementary Fig. S7).

Based on the levels and patterns of ADPRylation, these 8 cell lines together represented 2 of the patient groups that we defined in Fig. 1: groups 1 and 3 (Supplementary Fig. S4A and S4B). The 2 “group 1 similar” cell lines collectively exhibited greater than 50-fold sensitivity to Olaparib compared with the “group 3 similar” cell lines (Supplementary Fig. S8). Whether these group-specific effects hold and are observed for “group 2 similar” and “group 4 similar” cell lines will require further analysis of additional cell lines. Together, our data demonstrate that the levels and patterns of ADPRylation correlate with PARPi sensitivity in patient-derived cell lines. Furthermore, they suggest that ADPRylation levels may be a useful biomarker to predict clinical responses to PARPi.

In this study, we sought to identify a new biomarker that could predict responses to both conventional chemotherapy and PARPi in ovarian cancers. We focused on cellular ADPRylation, which we detected using reagents that we developed previously (22). We found that the levels and patterns of ADPRylation and PARP-1 protein segregated 34 high-grade serous ovarian cancers into distinct molecular phenotypes (Fig. 1), which exhibit dramatically different gene expression profiles (Fig. 3) and clinical outcomes in response to platinum-based chemotherapy (Fig. 4). Finally, we found that ADPRylation levels in patient-derived ovarian cancer cell lines correlate with sensitivity to the PARPi Olaparib independent of BRCA1/2 status (Fig. 5). Collectively, our results suggest that ADPRylation may be a useful biomarker for PARP inhibitor sensitivity in ovarian cancers.

The four ADPR-defined ovarian cancer groups exhibited distinct gene expression patterns. In future studies, it may be informative to compare our groupings to similar analyses based solely on gene expression (43). The gene expression analyses revealed cancer-relevant ontologies for the differentially expressed genes, including transcriptional regulation, cellular signaling, cell adhesion, extracellular organization, and inflammatory responses. Some of these gene ontologies are reminiscent of aggressive cancers, which rearrange their extracellular matrices to alter their ability to adhere and invade surrounding tissues (44, 45). Given that the gene expression patterns in a cell are indicative of the biology of the cell or tissue, we can infer that the levels and patterns of ADPRylation are also indicative of the biology of the cell or tissue. Although ADP-ribosylation is likely to be related to therapeutic responses to PARP inhibitors, it may also be more broadly indicative of the biology of the cancer (e.g., aggressive growth). This may make ADP-ribosylation broadly useful as a biomarker in cancers.

Although one goal of our analyses was to demonstrate that there is biological meaning for protein ADPRylation in a cancer context, we caution that converting this to a meaningful clinical test may be a difficult task. While our study has made headway in this regard, it is not our intention to imply that the groups we defined are necessarily the most clinically relevant way of segregating ovarian cancers. Such conclusions will require a much larger cohort of patient samples. We see our work as pointing the way to what might be possible, rather than an end-all, be-all set of conclusions about clinical states.

Finally, the correlations that we observed between ADPRylation and gene expression outcomes raise the possibility that ADPRylation may play a direct role in gene expression, Although PARP-1 and ADPRylation are known to play key roles in DNA damage surveillance and DNA repair, studies over the past 2 decades have implicated PARP-1 and ADPRylation in the regulation of chromatin structure and gene expression as well (15, 17).

Initial studies of ADPRylation in cancers, mainly PARylation, in cell lines and tumor specimens, have revealed significant variations in ADPRylation levels among different cancer types, as well across individual samples of a given cancer type (19, 20, 30, 46–48). Few studies, however, have investigated endogenous ADPRylation as a potential biomarker for responses to PARP inhibitor treatment. Oplustilova and colleagues observed a correlation between the levels of PARylation and responses to a PARP inhibitor (KU58948; ref. 19). They found that human cancer cell lines with the lowest levels of PAR (i.e., PC3, DU-145) had greater resistance to PARP inhibitor compared with cancer cell lines with higher levels of PAR (i.e., OVCAR 3, U2OS, CAPAN-1).

Similarly, in our studies, we observed the highest sensitivity to the PARP inhibitor Olaparib in patient-derived ovarian cancer cell lines with the highest basal levels of ADPRylation (e.g., OVCAR-3 and HCC5012). These cells resemble the tumor samples found in group 1, which had the highest levels of ADPRylation across all tumor samples. In addition, we observed resistance to Olaparib in patient-derived ovarian cancer cell lines with the lowest basal levels of ADPRylation (e.g., HCC5048 and HCC5024). These cells resemble the tumor samples found in group 3. Interestingly, the Olaparib responses that we observed in the 8 cell lines that we tested were independent of BRCA1/2 status. Given the gene expression outcomes described above, our results suggest that PARP-1 may be modulating PARP inhibitor sensitivity by activating or repressing specific target genes. Collectively, the results from published studies, as well as our results herein, support the idea that variations in ADPRylation may account for, or at least correlate with, variations in clinical responses to PARP inhibitors.

Our findings of elevated PARP inhibitor sensitivity in ovarian cancer cells with the highest levels of ADPRylation is supported by the work of Gottipati and colleagues (20), who observed hyperactivation of PARP-1 in HRR deficient cell lines and sensitivity to PARP inhibitor in cells with high levels of PAR. We did not observe any obvious relationship between ADPRylation levels and BRCA1/2 status in our study. However, we did not characterize HRR proficiency in our tumor specimens and cell lines beyond germline BRCA1/2 mutations. Thus, we cannot exclude the possibility that HRR deficiency is also playing a role in the clinical outcomes that we observed. The relationship between PARP, ADPRylation, and HRR proficiency should be investigated further.

W.L. Kraus and B.A. Gibson have ownership interest (including patents) in Patent No. 9599606: ADP-ribose detection reagents. W.L. Kraus is a founder and consultant for Ribon Therapeutics, Inc. J.S. Lea is a member of the advisory board for Clovis Oncology. No potential conflicts of interest were disclosed by the other authors.

The authors would like to dedicate this paper to A.F. Gazdar, who died in December 2018 during the completion of this work. A.F. Gazdar was a pre-eminent pathologist and world-renowned lung cancer expert, who shared valuable insights and discoveries of human cancers with the world.

Conception and design: L.B. Conrad, J.S. Lea, W. Lee Kraus

Development of methodology: L.B. Conrad, T. Nandu, J.S. Lea, W. Lee Kraus

Acquisition of data (acquired and managed patients, provided facilities, etc.): L.B. Conrad, K.Y. Lin

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): L.B. Conrad, T. Nandu, W. Lee Kraus

Writing, review, and/or revision of the manuscript: L.B. Conrad, T. Nandu, J.S. Lea, W. Lee Kraus

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): L.B. Conrad

Study supervision: J.S. Lea, W. Lee Kraus

Other (prepared reagents used in the manuscript): B.A. Gibson

The authors would like to thank Dr. Adi Gazdar, who generated and shared the human ovarian cancer cell lines prior to publication, and advised how best to grow and use them. The authors would also like to thank members of the Kraus lab and the Division of Gynecologic Oncology at UT Southwestern for critical comments and helpful suggestions on this work. This work was supported by a grant from the NIH/NIDDK (R01 DK069710) and funds from the Cecil H. and Ida Green Center for Reproductive Biology Sciences Endowment to W.L. Kraus.

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