Identification of inhibitors of ovarian cancer stem-like cells by high-throughput screening

Background Ovarian cancer stem cells are characterized by self-renewal capacity, ability to differentiate into distinct lineages, as well as higher invasiveness and resistance to many anticancer agents. Since they may be responsible for the recurrence of ovarian cancer after initial response to chemotherapy, development of new therapies targeting this special cellular subpopulation embedded within bulk ovarian cancers is warranted. Methods A high-throughput screening (HTS) campaign was performed with 825 compounds from the Mechanistic Set chemical library [Developmental Therapeutics Program (DTP)/National Cancer Institute (NCI)] against ovarian cancer stem-like cells (CSC) using a resazurin-based cell cytotoxicity assay. Identified sets of active compounds were projected onto self-organizing maps to identify their putative cellular response groups. Results From 793 screening compounds with evaluable data, 158 were found to have significant inhibitory effects on ovarian CSC. Computational analysis indicates that the majority of these compounds are associated with mitotic cellular responses. Conclusions Our HTS has uncovered a number of candidate compounds that may, after further testing, prove effective in targeting both ovarian CSC and their more differentiated progeny.


Background
Ovarian cancer is the most lethal of gynecological cancers [1] despite its typically high initial response rate to chemotherapy [2]. Recent evidence supports the existence of ovarian cancer stem-like cells (CSC), characterized by self-renewal capacity, ability to differentiate into distinct lineages, high invasiveness and resistance to a number of anticancer agents [3][4][5][6]. Since CSC have been shown to be resistant to most current chemotherapies, the frequent recurrence of ovarian cancer is believed, at least in part, to be attributable to the existence of chemo-resistant sub-populations of cancer cells embedded within bulk tumors [6]. For this reason, there is considerable current interest in the development of new chemotherapies that can effectively target this insidious subpopulation of tumor cells [7].
Thus far, searches for compounds that may be therapeutically effective against CSC have employed two alternative strategies. One approach has been to evaluate molecules known to be inhibitory against pathways believed to be deregulated in CSC (e.g., the Hedgehog, NOTCH, PTEN/AKT and WNT/β-catenin signaling pathways) [8]. This approach has resulted in the identification of several potential therapeutic agents that are currently in clinical trials [9]. A second approach is the high-throughput screening (HTS) of CSC-enriched cell populations with libraries of potential inhibitory compounds. This approach has been productively employed to identify candidate compounds displaying cytotoxic/ inhibitory effects on breast cancer [10] and glioma [11,12] CSC.
We have recently reported the isolation and characterization of ovarian CSC from an established ovarian cancer cell line OVCAR-3 [13]. These cells display a variety of features and molecular profiles characteristic of CSC previously isolated from ovarian and other cancer tissues [14][15][16][17]. Here we report the results of a high-throughput screening of 825 potential drugs (the National Cancer Institute's "Mechanistic Set" library) [18,19] against ovarian CSC and the subsequent identification of compounds that display significant potential for future development as ovarian CSC therapeutic agents.

Compounds
The NCI Mechanistic Set was provided by the Developmental Therapeutic Program (NCI/NIH) as a set of 825 compounds plated in eleven 96-well plates (plate numbers: 4520-4530; suffix: 69). Basic information on these compounds can be retrieved from the DTP website [20] using plate number as the search parameter. These compounds were selected from 37,836 compounds in the NCI repository to represent a broad range of growth inhibition patterns in the NCI 60 cell line screen [18,19] and consequently, they likely represent a diversity of the modes of action of these compounds. Compounds were supplied by DTP as 1 mM solutions in DMSO.

HTS and data analysis
Spheroids were dissociated to single cells using trypsin; trypsin was neutralized using Soybean Trypsin Inhibitor (Life Technologies, Grand Island, NY; Catalogue # 17075029), and cells were re-suspended in SCM to a density of 50,000 cells/mL. Cells were plated into flat bottom ultra-low attachment 96 well plates (Corning, Product #3474) in a volume of 198 μL per well (200 μL of SCM for blank wells) and incubated for 24 h at 37°C and humidified atmosphere with 5% CO 2 . Drug dilutions were prepared as follows: the eleven supplied NCI Mechanistic Set plates were copied (4 μL of DMSO solution per well) into sterile, round bottom polypropylene 96-well plates (Corning, Product #3359) and each drug was diluted with 22 μL media (working concentrations 153.8 μM). 3 μL of diluted library were added to 198 μL of cells (4 replicated wells for each drug), which resulted in final drug concentration of 2.29 μM. The plates were incubated for 96 h at 37°C and humidified atmosphere with 5% CO 2 . Thereafter, 20 μL of TOX8 reagent (Sigma-Aldrich) were added to each well and after 4-h incubation fluorescence intensities were measured for each well at 560 nm (excitation) and 590 nm (emission). The resazurin (Alamar blue)-based TOX8 reagent has been previously established as a reliable method for determining cell viability/cytotoxicity of tumor spheroid cell cultures [21,22].
Fluorescence intensities for replicated wells were analyzed using Grubb's test for detecting outliers at critical Z = 1.48 and outliers were removed from the dataset (test was applied only once for each replicated set of values). Percent growth of treated cells relative to untreated control cells was calculated from fluorescence intensities using the formula: %growth ¼ 100Â average for treated cells À average for blank wells ð Þ ½ = average for untreated cells ð Þ À ½ average for blank wells ð Þ Other parameters recommended by Inglese et al. [23], HTS plate design, assay performance evaluation and systematic error detection are presented in Additional file 1.
Of the 793 compounds that passed our assay performance standard, 99 displayed a single outlier among 4 replicated fluorescence intensity values and these outliers were removed before % growth values were calculated.
Statistical significance of differences between fluorescence intensities of drug-treated and untreated control wells was evaluated using Welch's t-test followed by Holm's step-down method for multiplicity adjustment [24]. Two-sided p-values were determined by Welch's t-test from raw fluorescence intensities corresponding to replicated treated and untreated control wells. These p-values represent probabilities that the difference between mean signal intensities for treated and control wells were obtained by chance. Holm's procedure was applied to 793 p-values (compounds that passed assay performance test) to counteract the problem of multiple comparisons and to ascertain that the probability of falsely identifying one or more compounds as significantly affecting growth of ovarian cancer stem-like cells is not more than 10%. Thus, compounds selected by controlling for family-wise error rate (FWER) of 0.10 were classified as compounds with statistically significant effect on growth of ovarian cancer stem-like cells. Drugs with % growth 80-110% were considered as compounds with no effect on growth of ovarian cancer stem-like cells.
Results of HTS were interpreted (i) in the context of the activity of FDA-approved drugs present in the library; (ii) in comparison with the potencies of screened compounds against OVCAR-3 cell line (parental cells from which CSC were isolated), and (iii) via mapping to self-organizing maps (SOMs).
A list of 97 FDA-approved oncology drugs was built from the data on the Approved Oncology Drug Set accessible at the DTP website [25] and made available in Additional file 2.
Potencies of library compounds against OVCAR-3 cell line (Dec 2010 release) were retrieved from the DTP website [26]. If multiple GI 50 values were available for the same compound, the average was taken without inclusion of default values (Note: NCI does not provide descriptive statistics (SD or SEM) for the determined GI 50 values for tested cells).
All 793 library compounds that passed our assay performance criteria and selected subsets of compounds active against CSC were mapped onto SOMs using the web-based tool 3D MIND developed by the Covell group at the National Cancer Institute [27]. The SOM method represents a type of artificial neural network trained using unsupervised learning to cluster high dimensional data and project them into a low dimensional space. SOMs used in this work was generated from GI 50 values for~30,000 compounds across 60 cell lines and consists of 1,350 hexagonal clusters with 9 major cellular response categories: mitosis (M), membrane function (N), nucleic acid metabolism (S), metabolic stress and cell survival (Q), kinases/phosphatases and oxidative stress (P) and 4 unexplored regions (RFJV) [28]. The significance of differences in the distribution of CSC active subsets and all library compounds to these areas was evaluated using Fisher's exact test with Yate's continuity correction and the difference was considered significant for two-sided p-value < 0.01.

Determination of GI 50
GI 50 values (concentrations of tested agents that inhibited growth of CSC cell cultures after 96-h incubation to 50% of the untreated control) were determined for 5 compounds (NSC72961, NSC302979, NSC82116, NSC673622 and NSC243928) randomly selected from those resulting in ≤ 50% cellular growth relative to untreated controls. GI 50 values were determined from concentration-response data generated by the same assay and cell system as used in our HTS. For each compound, 5 concentrations were used (15.6 nM, 62.5 nM, 250 nM, 1000 nM and 4,000 nM). Percent growth (control based normalization) was calculated as described in (Additional file 1: Table S1) and GI 50 values were determined by non-linear regression of logtransformed data using a normalized response-variable slope model (GraphPad Prism 5.01; GraphPad Software, Inc.) and expressed as mean ± SEM.

Results and discussion
HTS identified over 100 compounds that significantly inhibit ovarian CSC growth NCI's Developmental Therapeutics Program (DTP) maintains a repository of synthetic and naturally occurring potential anticancer drugs. A sub-set of these compounds, termed the "Mechanistic Set", represents the range of distinct growth inhibition patterns observed when the repository compounds were tested against the NCI-60 panel of cancer cell lines [18,19].
The CSC culture used in this study was derived from OVCAR-3 cells and is composed of tumor spheroids enriched with slowly proliferating self renewing stem-like cells that have been previously demonstrated to resist apoptosis after detachment from the surface (anoikis), a property known to be prerequisite for invasion and metastasis [13]. In addition, these cells have been shown to display significantly higher invasiveness, migration potential, and resistance to standard anticancer agents relative to the parental OVCAR-3 cells, as well as an ability to differentiate from a CD44-positive/ mesenchymal-like phenotype to CD44-negative/epithelial like phenotype [13].
CSC were seeded in 96-well plates and allowed to proliferate for 24 h prior to exposure to the 825 compounds comprising the NCI Mechanistic Set of experimental compounds (chemical library). After 96 h of exposure, % growth of treated cells relative to untreated controls was determined. To reduce the possibility of spurious results, we conducted an assay performance test and subsequently excluded 32 compounds from our survey that were associated with unreliable results (Additional file 3). Of the remaining 793 compounds (Additional file 4), 329 (41.5%) did not display an appreciable effect on ovarian CSC growth (80-110% growth), while 161 compounds (20.3%) displayed a statistically significant effect at FWER=0.10 with 158 compounds displaying an inhibiting effect (3.3-74.1% growth) and 3 compounds displaying a stimulatory effect (135.6-158.7% growth).
In order to focus on the most inhibitory compounds, we operationally defined CSC inhibitory compounds as those displaying a ≤ 50 % cell growth relative to untreated controls. Based on these criteria, 136 of the 793 compounds (17.2%) were classified as inhibitory (range of inhibitory growth: 3.3-50.4%) ( Table 1). These 136 inhibitory compounds are listed in Table 1 by a NCI compound number (NSC). A full list the various names associated with each NCI compound number is available as an ASCII file at the National Cancer Institute's Developmental Therapeutics Program website [29].
We randomly selected 5 of the 136 inhibitory compounds for confirmatory testing. In each case, the deter-  Included among the 793 compounds passing our performance test are 19 drugs previously approved by the FDA for cancer treatment ( Table 2). Five of these drugs are among the 136 compounds designated as CSC-inhibitory, but none are commonly used in the treatment of epithelial ovarian cancer. On the other hand, triethylenemelamine (altretamine, NSC9706), which is used for palliative treatment of persistent or recurrent ovarian cancer [30], induced non-significant stimulation of CSC growth in our screening ( Table 2).
The two most CSC-inhibitory of the FDA-approved drugs, dactinomycin (NSC 3035; 8.4% cell growth) [31] and plicamycin (mithramycin A, NSC 24559; 10.8% cell growth) [32] have both been previously reported to induce programmed cell death or apoptosis by inhibiting RNA transcription. Dactinomycin is used in the treatment of several cancers including gestational trophoblastic neoplasia [33] and Wilms' tumor [34]. Plicamycin has been used in the treatment of testicular cancer [35] and hypercalcemia associated with advanced malignancy [36]. The anti-microtubule drug vincristine (NSC24559) is used in the treatment of acute leukemias, Hodgkin lymphoma, and aggressive non-Hodgkin lymphoma, but is also included in combinations for treatment of small-cell lung cancer, breast cancers and some pediatric neoplasms [37]. Vinblastine (NSC49842) is used in the treatment of Hodgkin's lymphoma [37], and in combination with cisplatin and bleomycin in the treatment of testicular and ovarian germ cell cancers [38]. Mepacrine (quinacrine, NSC14229), an inhibitor of NFκB [39] and topoisomerase activity [40], is primarily used as an antimalarial drug [41]. In oncology, it is most commonly used for the treatment of pleural effusions in advanced malignant diseases [42].
Most ovarian CSC growth-inhibiting compounds also inhibit the growth of more differentiated ovarian cancer cells It has been established previously that normal stem cells are more resistant to the induction of apoptosis by radiation and cytotoxic agents than their differentiated progeny and similarly, CSC have been shown to display increased resistance to these same agents relative to the more differentiated cells that comprise the bulk of the tumor [43][44][45]. Indeed, it has been proposed that this dichotomy may contribute to the recurrence of cancer growth after the initial response of tumors to chemotherapeutic treatments [46]. Consistent with this view, we previously reported that several ovarian cancer drugs (e.g., NSC119875-cisplatin, NSC724770-docetaxel, NSC609699-topotecan) that are effective against ovarian cancer OVCAR-3 cells, are significantly less effective at inhibiting growth of ovarian CSC [13]. To assess if the apparent dichotomy in drug effectiveness between ovarian CSC and their more differentiated progeny is characteristic of the 136 ovarian CSC inhibitory compounds identified in our study, we sought to compare the results of our HTS with NCI's previous testing of compounds against the OVCAR-3 cell line. In the NCI program, GI 50 values (concentrations required to inhibit growth by 50%) on OVCAR-3 cells were determined for nearly all of the compounds used in our HTS. The OVCAR-3 GI 50 values (expressed as -log 10 GI 50 ) for 136 of the CSC-inhibitory compounds are presented in Table 1. The data output from our HTS is relative % growth rather than GI 50 values. However, since the concentration of compounds used in our HTS was 2.29 μM (see Methods), the GI 50 for compounds resulting in ≤ 50% growth of CSC is predicted to be ≤ 2.29 μM or ≥ 5.64 on the -log 10 scale (−log 10 2.29×10 -6 = 5.64). By comparing this value with the -log 10 GI 50 values previously determined for OVCAR-3 cells in the NCI study, we found that 73% (99/136) of the compounds that we designated as CSC inhibitory compounds are also inhibitory for OVCAR-3 cells (i.e., -log 10 GI 50 ≥ 5.64). This suggests that there may be a number of inhibitory compounds with the potential to target both ovarian CSC and their more differentiated progeny. Of the remaining 37 (136-99) CSC-inhibitory compounds, 5 were not previously tested on OVCAR-3. Thus, based on our criteria, we classified 32 compounds to be preferentially inhibitory for CSC (Table 3). Among the 99 compounds found to be co-inhibitory for CSC and OVAR-3 cells, four have previously been FDA approved for cancer treatment (NSC3053-dactinomycin; NSC24559-plicamycin; NSC49842-vinblastine; NSC67574vincristine) ( Table 2). Only one of the previously approved cancer drugs, the NFκB-inhibitor Mepacrine (NSC14229),  is included among those compounds classified as preferentially inhibitory for ovarian CSC (37.6% growth).

Computational model classifies CSC inhibitory compounds into predicted cellular response groups
Neither the mode of action nor the molecular targets of the 32 compounds classified by us as preferentially inhibitory for CSC have, as yet, been definitively determined. However, there are a variety of computational tools that can be informative in predicting the putative cellular responses to these compounds and in suggesting lines of future investigation. One such tool, developed by the Covell group at NCI [27,28,47], utilizes data from the treatment of 60 representative human cancer cell lines (the NCI-60 panel) [48] with nearly 30,000 compounds including those investigated    in our HTS. These data were used to generate a selforganizing map (SOM) that visualizes the position of 30,000 screened compounds within 9 major cellular response categories: mitosis (M), membrane function (N), nucleic acid metabolism (S), metabolic stress and cell survival (Q), kinases/phosphatases and oxidative stress (P) and 4 unexplored regions (RFJV).
Using this SOM, we were able to visually compare the predicted cellular responses of all of our screened compounds (that passed assay performance criteria) relative to those identified as having an inhibitory effect on CSC. Shown in Figure 3 is a SOM upon which we have mapped (a) all 793 compounds used in our HTS; (b) the 136 compounds identified as inhibitors of CSC (including compounds that also inhibit OVCAR-3 cells); (c) the 99 compounds identified as co-inhibitory of CSC and OVCAR-3 cells; and (d) the 32 compounds that exert a CSC-specific inhibitory effect.
The results indicate that compared to all 793 evaluable compounds, the 136 identified as inhibitory for CSC are highly significantly enriched for compounds associated with M (mitotic) cellular responses (M: 36.1% vs 18.7%; p<0.0001; Figure 3b and Table 4). This region of SOM contains many compounds known to interfere with microtubule and/or actin filaments, such as taxanes, derivatives of colchicine, vinca alkaloids, rhizoxin and nocodazole [47]. This region also contains compounds associated with inhibition of the DNA polymerase pathway and is associated with the Gene Ontology (GO) terms: Mitotic checkpoint, Cytokinesis, DNA topological change, Cell cycle (Biological Processes); Nucleus, Kinetochore (Cellular Components), and DNA topoisomerase activity (Molecular Function) [28].
The 99 compounds that were co-inhibitory for OVCAR-3 and CSC also displayed a significant enrichment for M cellular responses (M: 38.5% vs 18.7%; p<0.0001; Figure 3c and Table 4). When the 32 compounds classified as preferentially inhibitory for CSC were mapped, the enrichment for M (mitotic) cellular responses was no longer apparent (Figure 3d). Although this may be attributable to the fact that CSC are less mitotically active than cancer epithelial cells [43], the relatively small number of compounds (32) in this category precludes definitive conclusions.