Small-molecule inhibitors of PARP1/2, such as olaparib, have been proposed to serve as a synthetic lethal therapy for cancers that harbor BRCA1 or BRCA2 mutations. Indeed, in clinical trials, PARP1/2 inhibitors elicit sustained antitumor responses in patients with germline BRCA gene mutations. In hypothesizing that additional genetic determinants might direct use of these drugs, we conducted a genome-wide synthetic lethal screen for candidate olaparib sensitivity genes. In support of this hypothesis, the set of identified genes included known determinants of olaparib sensitivity, such as BRCA1, RAD51, and Fanconi's anemia susceptibility genes. In addition, the set included genes implicated in established networks of DNA repair, DNA cohesion, and chromatin remodeling, none of which were known previously to confer sensitivity to PARP1/2 inhibition. Notably, integration of the list of candidate sensitivity genes with data from tumor DNA sequencing studies identified CDK12 deficiency as a clinically relevant biomarker of PARP1/2 inhibitor sensitivity. In models of high-grade serous ovarian cancer (HGS-OVCa), CDK12 attenuation was sufficient to confer sensitivity to PARP1/2 inhibition, suppression of DNA repair via homologous recombination, and reduced expression of BRCA1. As one of only nine genes known to be significantly mutated in HGS-OVCa, CDK12 has properties that should confirm interest in its use as a biomarker, particularly in ongoing clinical trials of PARP1/2 inhibitors and other agents that trigger replication fork arrest. Cancer Res; 74(1); 287–97. ©2013 AACR.
Enzymes within the PARP superfamily catalyze the polymerization of PAR chains on substrate proteins, using β-NAD+ as a cofactor (PARylation). The best-studied member of this superfamily, PARP1, has a relatively well-defined role in the DNA damage response (DDR) and binds damaged DNA, PARylates DNA repair effectors such as XRCC1 and conducts autoPARylation, an event that drives its release from DNA (1).
Small-molecule PARP1 inhibitors, which also inhibit the homologous PARP2 enzyme (hereafter referred to as PARP1/2 inhibitors), were originally developed as chemo- and radiosensitizing agents. These agents also elicit synthetic lethality in BRCA1- or BRCA2-deficient cells, an observation that led to their clinical assessment in patients with germline BRCA1 or BRCA2 mutations (2). In phase 1 and 2 trials, olaparib single-agent treatment has delivered significant and sustained antitumor responses in patients with germline BRCA mutations, without causing many of the side effects associated with standard chemotherapies. Furthermore, when used as a maintenance therapy after carboplatin treatment in patients with high-grade serous ovarian cancer (HGS-OVCa), a disease characterized by a relatively high frequency of familial and somatic BRCA mutations, olaparib (KuDos/AstraZeneca) has been shown to significantly improve progression-free survival (PFS) when compared with a placebo and in patients with BRCA1- or BRCA2-mutant tumors and can reduce the time to recurrence by approximately 85% (3, 4).
So far, the genetic dissection of PARP1/2 inhibitor sensitivity has relied upon the study of single genes involved in DNA repair (5) or small subsets of genes/proteins, such as the kinome (6). To identify novel genetic determinants of PARP1/2 inhibitor sensitivity that could have an impact on the clinical development of these agents, we conducted a genome-wide PARP1/2 inhibitor synthetic lethal screen. The results of this screen confirm the impact of the DDR on the cellular response to PARP1/2 inhibition and identify novel genetic determinants, such as CDK12, that could be used as candidate predictive biomarkers in both existing and future clinical trials.
Materials and Methods
The PARP inhibitor AZD2281/Olaparib was obtained from Selleck Chemical, and paclitaxel and cisplatin were obtained from Sigma Aldrich. Drugs were used as previously described (7–9).
Short hairpin RNA screen
We used the OpenBiosystems GIPZ human short hairpin RNA (shRNA) library. Viral pools encompassing about 9,500 shRNAs were used to transduce MCF7 cells with a final representation of about 1,000 cells per shRNA construct (Supplementary Fig. S1C). Seventy-two hours after infection, cells were divided into cohorts and exposed to either 1.5 μmol/L olaparib or dimethyl sulfoxide (DMSO) for 14 days. shRNA representation in the surviving cell populations was estimated via massively parallel sequencing as described in Supplementary Materials and Methods.
PEO1, PEO14, OV56, COV318, COV504, and SKOV3 were obtained from European Collection of Cell Cultures. PEO1 and PEO14 cells were grown in RPMI-1640 containing 2 mmol/L glutamine, 2 mmol/L sodium pyruvate, and 10% (v/v) FBS. OV56 cells were grown in Dulbecco's Modified Eagle's Medium (DMEM):F12 Nutrient Mixture containing 2 mmol/L glutamine, 5% (v/v) FBS, 0.5 μg/mL hydrocortisone (Sigma Aldrich), and 10 μg/mL insulin (Sigma Aldrich). COV318 and COV504 cells were grown in DMEM containing 2 mmol/L glutamine and 10% (v/v) FBS. SKOV3 cells were grown in McCoy's 5A media containing 2 mmol/L glutamine and 15% (v/v) FBS. OV90 and CAOV3 were obtained from American Type Culture Collection. OV90 cells were grown in a 1:1 mixture of MCDB105 containing 1.5 g/L sodium bicarbonate and Medium 199 containing 2.2 g/L sodium bicarbonate and 15% (v/v) FBS. CAOV3 cells were grown in DMEM containing 2 mmol/L glutamine and 10% (v/v) FBS. The identity of cell lines was confirmed by short tandem repeat (STR) typing using the StemElite Kit (Promega) in February 2013. Cell lines were confirmed as being mycoplasma negative using the MycoAlert Kit (Lonza) routinely throughout experimentation. Cell lines were transfected with SMARTpool siRNAs obtained from Dharmacon or Santa Cruz Biotech using RNAiMax (Invitrogen) transfection reagent (see also Supplementary Materials and Methods).
Cells were lysed in NP250 buffer (20 mmol/L Tris, pH 7.6, 1 mmol/L EDTA, 0.5% NP40, 250 mmol/L NaCl) containing protease inhibitor cocktail tablets (Roche). Protein quantification was estimated using BioRad Protein Assay Reagent (BioRad). Whole-cell lysates were electrophoresed on Novex 4% to 12% gradient bis–tris pre-cast gels (Invitrogen) and immunoblotted overnight at 4°C with antibodies listed in Supplementary Table S1.
Homologous recombination assay
A synthetic repair reporter was used as previously described (8, 10–12). HeLa cells harboring a single-copy genomic integration of the DR-GFP reporter were transfected with siRNA targeting CDK12, BRCA1, BRCA2, or an siControl. Twenty-four hours later, cells were transfected with the I-SceI expression vector, pcBASce, or an empty vector control (11). Forty-eight hours later, homologous recombination (HR) frequency was estimated by quantifying the frequency of GFP-positive cells using fluorescence-activated cell sorting (FACS; ref. 11).
The quantification of nuclear RAD51 foci was conducted as described previously (7). See also Supplementary Materials and Methods for details.
Quantitative real-time PCR
Quantitative real-time (RT)-PCR was carried out using Assay-on-Demand primer/probe sets (Applied Biosystems). Gene expression was calculated relative to the expression of GAPDH and adjusted relative to expression in nontargeting shRNA (shNTC)-infected cells.
Overall survival (OS) analysis was conducted using stage II–IV HGS-OVCa samples from The Cancer Genome Atlas (TCGA) dataset (13–15) as described in the Supplementary Materials and Methods.
In vivo efficacy studies
In vivo efficacy study was conducted using OV90 cells infected with either shNTC or shCDK12 expression constructs. A detailed protocol is described in Supplementary Materials and Methods.
A genome-wide synthetic lethality screen for sensitivity to a clinical PARP1/2 inhibitor
To identify determinants of PARP1/2 inhibitor response, we designed a high-throughput RNA interference (RNAi) synthetic lethality screen (Supplementary Fig. S1A). For this screen, we selected MCF7 breast tumor cells as they are BRCA1 and BRCA2 wild-type, olaparib-resistant (Supplementary Fig. S1B) and amenable to high-efficiency lentiviral infection (Supplementary Fig. S1C). Cells were infected with a GIPZ human shRNA library encompassing 57,540 lentiviral shRNA expression constructs targeting 16,487 unique human protein-coding genes (Ensembl 56). The shRNA library was divided into 6 pools, with each pool encompassing between 9,541 and 9,601 shRNAs. MCF7 cells were infected with each GIPZ pool at a multiplicity of infection of 0.8, generating about 1,000 infected cells per shRNA construct. This ratio of cells infected per shRNA construct was maintained throughout the experiment (Supplementary Fig. S1C). Seventy-two hours after infection, cell populations were divided into 2 cohorts, one exposed to the clinical PARP1/2 inhibitor olaparib (KuDOS/AstraZeneca) and the other exposed to DMSO, the olaparib vehicle. To maximize the potential for identifying synthetic lethal/sensitizing effects, cells were exposed to 1.5 μmol/L olaparib, a concentration that caused a 20% reduction in cell survival (surviving fraction 80, SF80; Supplementary Fig. S1B). In total, cells were exposed to olaparib or DMSO for 14 days (10 population doublings) so as to model chronic olaparib exposure as used in the clinic.
To identify shRNA constructs that modulated the response to olaparib, we estimated shRNA enrichment and depletion in cells that survived olaparib or DMSO exposure using massively parallel sequencing. In brief, gDNA from surviving cell populations was recovered and shRNA target sequences were PCR-amplified using primer sequences compatible with Illumina GAIIx sequencing (see Supplementary Materials and Methods). PCR amplicons were sequenced on a Illumina GAIIx platform, generating more than 1,000 short reads for each shRNA in the library. After aligning short-read sequences to the known sequence of shRNAs present in each pool, we used the frequency of each short-read to estimate the frequency of shRNAs in each surviving population. After data normalization to account for variation between viral pools and replicas (see Supplementary Materials and Methods), we identified shRNA constructs that modulated the MCF7 response to olaparib by comparing the abundance of shRNA-specific short-reads in olaparib and vehicle-treated cultures. Overrepresentation of shRNA sequence in an olaparib-treated sample indicated a resistance-causing effect, whereas underrepresentation indicated a sensitization effect. These over- and underrepresentation effects were quantified as drug effect (DE) Z or standardized scores (See Supplementary Materials and Methods), with negative DE scores indicating sensitization effects and positive DE scores indicating resistance-causing effects (Fig. 1A–C). In total, we conducted 2 biologic replicas of the screen and calculated average DE scores for each shRNA from the replica data (Fig. 1A).
We calculated DE Z score values using the median absolute deviation (MAD) and defined sensitization effects as those shRNAs that gave DE Z scores <−1.96, a threshold approximately equivalent to P < 0.05 and likewise resistant effects as those shRNAs that gave Z score >1.96 (Fig. 1B, Supplementary Tables S2 and S3). In total, we identified 2,339 shRNA constructs targeting 2,208 different genes causing an enhancement of olaparib sensitivity in MCF7 cells and 2,043 shRNA constructs targeting 1,902 genes causing an increase in olaparib resistance. We found that of 2,208 different genes appearing in the candidate sensitivity list, the well-established PARP1/2 inhibitor sensitivity gene BRCA1 (9 shRNAs with DE Z score <−1.96) and the ATAD5 gene (ATPase family, AAA domain containing 5, 4 shRNAs with DE Z score <−1.96), were represented by more than 2 shRNAs with DE Z scores <−1.96. ATAD5 is known to modulate proliferating cell nuclear antigen (PCNA) deubiquitination and genomic stability (16–18), and we found that silencing of ATAD5 enhanced olaparib sensitivity and reduced RAD51 foci formation (a major determinant of PARP1/2 inhibitor sensitivity; Supplementary Fig. S2), giving us some confidence in the results of the screen. Five percent of the genes (119 genes) in the candidate sensitivity list were represented by two shRNAs with DE <−1.96, whereas the vast majority of genes in the sensitivity list were represented by one shRNA with DE <−1.96 (2,089 genes, 95%). Similarly, in the candidate resistance-causing gene list, there were 141 genes represented by 2 shRNAs (7%) and 1,761 genes (93%) represented by one shRNA. From our experience with this shRNA library (19, 20), the predominance of hits represented by only one shRNA is common and explains the rationale behind efforts that many have made to increase the number of shRNA species per gene in newer-generation shRNA libraries (21).
Using pathway annotation tools such as KEGG and literature sources, we annotated the list of candidate “hit” genes to identify the predominant known molecular networks explaining the response to olaparib. To facilitate network analysis, we annotated all genes in which at least one shRNA caused either sensitivity or resistance as defined above. In the case of the olaparib sensitivity-causing effects, the 2 most predominant networks represented were DNA repair processes and DNA cohesion/chromatin remodeling. For example, we found evidence for an involvement in DNA repair for at least 74 genes in the candidate sensitivity-causing gene list (Fig. 1D, Supplementary Fig. S3 and Supplementary Table S4). Of these, 65% (48 genes) have already been implicated in either BRCA1 function, DNA double-stand break repair, the Fanconi's anemia pathway, or HR (Fig. 1D and E, Supplementary Fig. S3 and Supplementary Table S4). Furthermore, 14 genes in the candidate sensitivity gene list have previously been reported as modulating PARP1/2 inhibitor sensitivity (BRCA1, NBN, FANCD2, FANCC, RAD51, MCM3, CDK7, GTF2H3, LIG3, POLH, PTEN, USP1, RAD51D, and RAD51C; Supplementary Table S4; refs. 6, 8, 22–25). The identification of a series of HR genes and known PARP1/2 inhibitor sensitization effects in our candidate hit list gave us considerable confidence in the screen data and supports the hypothesis that HR is a key determinant of PARP1/2 inhibitor sensitivity. However, we did note that some critical HR genes were not present in our candidate hit list, such as BRCA2 and DSS1; it seems possible that the shRNA constructs for these genes do not cause a level of gene silencing sufficient to elicit HR deficiency, reinforcing the idea that the negative predictive value of such screens (i.e., the ability to identify genes that do not cause the phenotype of interest) might be limited.
We also noted that a number of cohesins, chromatin remodeling proteins, and replication-associated genes that are involved in DNA repair were found in our candidate sensitization gene list (e.g., the cohesion-associated genes, RAD21, ESCO1, ESCO2, and SMC3; the MCM protein-coding genes MCM2, 3, and 6; and the topoisomerase-coding genes TOP3A and TOP2B; Fig. 1F; Supplementary Table S4 and Supplementary Fig. S3) highlighting the potential impact of sister chromatid cohesion and chromatin remodeling in the response to PARP1/2 inhibitor–induced lesions. This latter observation is consistent with work identifying cohesins and associated proteins as being critical to the DDR (10, 26). In addition, a number of other DNA repair proteins not so far linked to HR were also identified in the candidate sensitization gene list, including those involved in base excision repair (BER) and nucleotide excision repair (NER; Supplementary Table S4), such as LIG3 and POLB (for the purposes of classification, defined as BER proteins; ref. 27) and GTF2H3 and CDK7 (both NER). Of these, LIG3, GTF2H3, and CDK7 have previously been implicated in PARP1/2 inhibitor sensitivity (23). The identification of multiple components of DDR not implicated in HR in the candidate gene list here suggests that the response to PARP1/2 inhibitor–driven DNA lesions is perhaps more complex.
Among the olaparib resistance-causing effects, we noted that 2 shRNAs targeting the olaparib target PARP1 caused olaparib resistance with DE Z scores of 2.0 and 3.4 (Supplementary Table S3). This observation is consistent with recent work identifying PARP1 deficiency as a cause of PARP inhibitor resistance in cells with functional HR (28, 29). It is suggested that the cytotoxic DNA lesions in cells exposed to PARP1/2 inhibitors are trapped PARP1–DNA complexes (28). These lesions most likely form as catalytic PARP inhibitors reduce PARP1 autoPARylation, an event that is normally required for the release of PARP1 from DNA. In the absence of PARP1, these DNA lesions do not form and the cytotoxicity normally caused by PARP1/2 inhibitors is reduced (28, 29).
Integration of the olaparib sensitization profile with tumor mutation data
One of our aims in this study was to identify genetic biomarkers that could be used to inform the clinical use of PARP1/2 inhibitors. Already, a number of cancer-associated mutations predict sensitivity to PARP1/2 inhibitors, such as loss-of-function mutations in BRCA1 and BRCA2. To identify other loss-of-function cancer mutations that could predict sensitivity to olaparib, we integrated our list of candidate shRNA sensitivity effects with publically available data describing somatic and germline mutations in a range of cancer histologies (30–32). As predicted, this analysis highlighted the presence of recurrent cancer-associated mutations in known olaparib sensitization genes identified in our shRNA screen such as BRCA1, ATM, FANCD2, FANCE, and PTEN. CDK12 (ENSG00000167258, also known as CRK7, CRKR, CRKRS), a gene represented in the olaparib shRNA sensitization list (DE = −4.3; Supplementary Table S2) was identified as 1 of only 9 genes found to be significantly mutated in HGS-OVCa (13). In a recent genomic analysis of 316 HGS-OV tumors, 9 tumors were found to have somatic CDK12 mutations (3%); in the same dataset, BRCA1 and BRCA2 were found to be mutated in 12% (BRCA1) and 11% (BRCA2) of tumors (13). Five of the CDK12 mutations in HGS-OVCa were nonsense or insertion/deletion mutations likely to be loss-of-function mutations (Fig. 2A and Supplementary Table S5). Furthermore, as Spellman and colleagues noted, the additional 4 missense CDK12 mutations identified in HGS-OVCa clustered in the protein kinase coding domain of CDK12, suggesting that they could compromise kinase activity (13). A subsequent analysis of the variant allele frequency of CDK12 mutations in HGS-OVCa has also suggested that loss of the wild-type CDK12 allele occurs in the majority of CDK12-mutant tumors, further arguing the case for CDK12 as an HGS-OVCa tumor suppressor gene (33). By analyzing data from the TCGA study describing the mutational spectrum of HGS-OVCa (13), we found that 95% of tumor BRCA1 mutations (36 of 38 cases) were mutually exclusive with mutations in BRCA2, an expected observation given the involvement of these 2 genes in a common pathway (HR; ref. 13). We also found that the majority of tumor CDK12 mutations in this same data set (78%, 7 of 9 cases) were mutually exclusive with mutations in either BRCA1 or BRCA2 (Fig. 2B). Although the mutual exclusivity relationships between BRCA1, BRCA2, and CDK12 mutations were not complete (i.e., 100%), the distribution of mutations in these 3 genes across the TCGA dataset may represent an alternative means to confer the same phenotype, suggesting epistasis or a common pathway being compromised.
CDK12 and PARP1/2 inhibitor sensitivity
The appearance of CDK12 in our olaparib sensitization gene list, alongside the CDK12 mutational data, suggested the possibility that loss of CDK12 function could sensitize tumor cells to PARP1/2 inhibitors and that loss of CDK12 function in HGS-OVCa could be a predictive biomarker for response to this developmental class of agents.
To test this, we first assessed the olaparib sensitivity of a panel of HGS-OVCa cell line models using the gold-standard clonogenic assay format, exposing cells to olaparib for 2 weeks. This analysis showed a gradation of response from the sensitive PEO1 (hemizygous BRCA2 p.Y1655X mutant; ref. 34) model through to the resistant SKOV3 cell line (Fig. 3A and Table 1). Next generation sequencing of most of these models has not identified CDK12 mutations (35), however, by carrying out RT-PCR and Western blot analysis of CDK12 across the tumor cell line panel (Fig. 3B and C), we found a significant correlation between reduced CDK12 expression and olaparib sensitivity (Fig. 3B and D, P < 0.05; Student t test).
|Cell line .||SF50, μmol/L .|
|Cell line .||SF50, μmol/L .|
To directly assess the relationship between CDK12 deficiency and PARP1/2 inhibitor sensitivity, we silenced CDK12 in the profoundly olaparib-resistant OV90 ovarian cancer cell line and assessed olaparib sensitivity. Three different CDK12 shRNA constructs caused partial silencing of CDK12 (Fig. 4A and B) and a modest yet significant increase in olaparib sensitivity (Fig. 4C, P < 0.05, ANOVA), at a scale equivalent to that achieved by a shRNA construct targeting BRCA1. In addition, an siRNA targeting CDK12 also sensitized OV90 cells to olaparib (Fig. 4D, P < 0.001, ANOVA). shRNA targeting CDK12 also sensitized an additional HGS-OVCa model, PEO14, to olaparib (Supplementary Fig. S4A and S4B, P < 0.05, ANOVA), suggesting the generality of this effect. We analyzed whether restoring CDK12 expression could confer PARP1/2 inhibitor resistance in a tumor cell line with low levels of endogenous CDK12 expression, such as CAOV3. Ectopic expression of CDK12 significantly restored olaparib resistance [Supplementary Fig. S4C, P < 0.05, ANOVA for CDK12 cDNA expression construct vs. a control (empty vector) cDNA expression construct]. Sensitivity to the microtubule poison paclitaxel was not increased by CDK12 silencing (Supplementary Fig. S4D), suggesting that CDK12 dysfunction does not in general confer sensitivity to cancer drugs.
We also assessed the hypothesis that the relationship between CDK12 dysfunction and PARP1/2 inhibitor sensitivity could be explained by a reduction of HR, the predominant pathway identified in our PARP1/2 inhibitor synthetic lethality screen (Supplementary Fig. S3). To assess this, we measured HR activity using a synthetic DNA HR substrate (DR-GFP) that generates a GFP signal once an experimentally induced DNA double-strand break (DSB) has been repaired by HR (11, 12). HeLa cells harboring the DR-GFP reporter system were reverse-transfected with siRNA targeting CDK12, BRCA1, or BRCA2 and 24 hours later, a single DSB was introduced in the synthetic DNA substrate via the expression of the I-Sce1 restriction endonuclease (11). CDK12 silencing caused a 70% reduction in HR, a level of suppression similar to that caused by siRNA targeting BRCA1 or BRCA2 (Fig. 5A, P < 0.0001, Student t test). We also assessed another marker of HR, the ability to form nuclear RAD51 foci in response to DNA damage (5). Silencing of CDK12 resulted in a significant reduction in ionizing radiation–induced RAD51 foci formation, when compared with control-treated cells (Fig. 5B and C, P < 0.001, Student t test), an observation consistent with the DR-GFP data and supporting the hypothesis that the modulation of PARP1/2 inhibitor sensitivity via reduced CDK12 activity is mediated by a defect in HR.
Blazek and colleagues recently reported that CDK12 dysfunction results in a reduction in expression of key DNA repair proteins such as BRCA1, FANCI, FANCD2, and ATR, as well as sensitivity to chemotherapies that stall replication forks such as mitomycin C and camptothecin (36). We found that CDK12 silencing in 2 HGS-OVCa models also caused sensitivity to the replication fork stalling agent cisplatin (P < 0.001, ANOVA for each CDK12 shRNA vs. shNTC; Supplementary Fig. S4E and S4F). In the first instance, we tested whether mRNA transcript levels of BRCA1, FANCD2, ATR, and FANCI were regulated by CDK12 in HGS-OVCa cells infected with CDK12 shRNAs (Supplementary Fig. S5A). We noted that levels of BRCA1 mRNA and protein were reduced by 2 different CDK12 shRNA expression constructs, whereas levels of other DNA repair proteins, such as ATR, were not (Fig. 5D and E and Supplementary Fig. S5A and B, P < 0.05, Student t test). The suppression of this key HR gene (BRCA1) provides a possible mechanistic explanation for the HR dysfunction and PARP1/2 inhibitor sensitivity seen in CDK12-defective cells. As HR and PARP1/2 inhibitor resistance can be restored in BRCA1-deficient cells by loss of 53BP1 activity (37–39) and BRCA1 levels were reduced by CDK12 suppression (Fig. 5D and E), we tested the possibility that inhibition of 53BP1 could also restore PARP1/2 inhibitor resistance in the absence of CDK12 expression. Gene silencing of 53BP1 in combination with silencing of CDK12 (Supplementary Fig. S5C) caused a modest increase in olaparib resistance (Fig. 5F, P < 0.0001, Student t test), consistent with the hypothesis that the loss of BRCA1 expression could play some role in mediating PARP1/2 inhibitor sensitivity in HGS-OVCa tumor cells with low CDK12. CCNK (also known as cyclin K) is the cyclin partner of CDK12 and regulates CDK12 expression (36). Using a previously validated RNA interference reagent for CCNK (36), we established that targeting of CCNK also caused olaparib and cisplatin sensitization in a HGS-OVCa model, at a scale equivalent to that achieved by CDK12 silencing (Supplementary Fig. S6A and S6B).
Given the data suggesting the role of CDK12 in HR, we sought to determine whether the extent of tumor CDK12 mRNA expression correlated with response to platinum therapy. Using survival data from patients treated with platinum therapy in the recent TCGA study, in which corresponding tumor mRNA levels were also quantified (13), we found that the patients with the lowest levels of tumor CDK12 mRNA expression had improved OS compared with those with elevated levels of CDK12 mRNA [Fig. 6A, P = 0.0076 (log-rank test), HR = 0.55; and Fig. 6B, P = 0.0375 (Student t test)]. Of note, the patients with tumor CDK12 mutations predicted to be most deleterious to CDK12 function (i.e., truncating mutations such p.L122fs, p.Q602*, p.W719*, and p.L926fs) fell into the CDK12 “low” cohort in this analysis.
On the basis of this preliminary clinical data, we assessed the therapeutic effect of olaparib on CDK12-deficient tumor cells in vivo. OV90 cells stably expressing a CDK12 shRNA expression construct (shCDK12+) or a control, nontargeting (shNTC+) shRNA construct were subcutaneously xenografted into female athymic nude mice. Once tumors had established and were measurable, mice were randomized into 1 of 4 treatment groups: (i) an shCDK12+ OV90 xenograft-bearing cohort treated with vehicle, (ii) an shCDK12+ OV90 xenograft-bearing cohort treated with olaparib, (iii) an shNTC+ OV90 xenograft-bearing cohort treated with vehicle, and (iv) an shNTC+ OV90 xenograft-bearing cohort treated with olaparib. There was a clear and statistically significant inhibitory effect of olaparib on tumor volume in shCDK12+ tumor xenografts compared with the other cohorts (Fig. 6C and D, P < 0.001, ANOVA for the shCDK12+ olaparib-treated cohort vs. each other cohort; all other comparisons returned a nonsignificant P value). The treatment regime was well-tolerated, with none of the mice showing a significant change in body weight (Supplementary Fig. S6).
As far as we are aware, this is the first attempt to define on a genome-wide scale the compendium of genes that controls the cellular response to a clinical PARP1/2 inhibitor. The initial analysis of this data confirms that the status of the DDR apparatus and particularly, HR is a key factor in determining the cytotoxicity of olaparib. Furthermore, genes that control chromatin remodeling and sister chromatid cohesion, perhaps as part of the DDR, also seem to modulate the response to olaparib. Subsequent interrogation of the compendium of candidate PARP1/2 inhibitor genes generated in this study could in future identify additional molecular networks that explain how cells respond to these drugs.
One of the major issues in the clinical development of PARP1/2 inhibitors is the identification of biomarkers other than BRCA1 and BRCA2 gene mutations that predict a favorable response to therapy. For example, phase 1 and 2 studies established that olaparib could be administered safely and that it showed significant responses in germline BRCA mutation carriers with breast, ovarian, or prostate tumors. However, not all trials have been so positive; efforts to target patient subgroups without germline BRCA mutations, but which might have similar HR defects such as triple-negative breast cancers, have not met with similar success (40). Similarly, the mixed responses to olaparib in patients with HGS-OVCa (4), a patient subgroup selected for trial because of its presumed high frequency of tumor-specific HR defects, argues the case for the identification of additional biomarkers that could be used to direct the use of PARP1/2 inhibitors. We propose that CDK12 should be considered as a candidate PARP1/2 inhibitor response biomarker and assessed in additional patient material from existing PARP1/2 inhibitor trials and integrated into the design of future trials, especially given its recurrent mutation in HGS-OVCa.
One future use of the compendium of genes described here is in the development of metagenes that predict the response to therapy. For example, we, and others, have previously shown that aggregate measures of gene expression from groups or modules of genes (metagenes) identified in functional profiling studies can correlate with patient responses (19). Once the full panoply of genomic and transcriptomic profiling of tumors from patients on PARP1/2 inhibitor trials becomes available, such analyses may become possible. Of course, such an approach should take into account the limitations of such in vitro genetic screens. In this particular case, we note that the screen was carried out in one tumor cell model and that cellular responses (and the genes that control these responses) may differ from cell line to cell line. Furthermore, the shRNA library used here could be improved by increasing the number of different shRNA constructs per gene and/or improving the ability of each construct to silence its target gene; both of these improvements could decrease the false-positive rate and enhance the overall quality of the data. Carrying out genetic screens such as this in an in vivo setting could also enhance the modeling of the tumor response within a more physiologic setting.
Disclosure of Potential Conflicts of Interest
C.J. Lord has ownership interest (including patents) in Institute of Cancer Research. A. Ashworth has expert testimony and may benefit financially from the development of PARP inhibitors through patents held jointly with AstraZeneca through the Institute of Cancer Research "rewards to inventors" scheme. No potential conflicts of interest were disclosed by the other authors.
Conception and design: I. Bajrami, J.R. Frankum, C.J. Lord, A. Ashworth
Development of methodology: I. Bajrami, J.R. Frankum, F. Rehman, D. Sims, C.J. Lord
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): I. Bajrami, J.R. Frankum, A. Konde, F. Rehman, R. Rafiq, I. Kozarewa, I. Assiotis, K. Fenwick, R. Natrajan, C.J. Lord
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): I. Bajrami, J.R. Frankum, J. Campbell, D. Sims, L. Chen, C.J. Lord
Writing, review, and/or revision of the manuscript: I. Bajrami, J.R. Frankum, D. Sims, C.J. Lord, A. Ashworth
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): I. Bajrami, J.R. Frankum, A. Konde, R.E. Miller, R. Rafiq, S. Hooper, C.J. Lord
Study supervision: I. Bajrami, J.R. Frankum, R. Brough, C.J. Lord
This work was funded by Breakthrough Breast Cancer (CTR-Q3-Y1), Cancer Research UK (C347/A8363), the European Union [EUROCAN (HEALTH-FP7-2010-260791) and DDR projects (HEALTH-FP7-2010-25989)]. Alan Ashworth is supported by a Stand Up To Cancer Dream Team Translational Cancer Research Grant, a Program of the Entertainment Industry Foundation (SU2C-1559-G-MD204).
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.