Materials and Methods

Materials.

Compounds used for this study were obtained through the DTP of the NCI (Bethesda, MD). PhA was obtained from Frontier Scientific (Logan, UT). FTC was prepared by Thomas McCloud (Screening Technologies Branch, DTP, NCI). Topotecan was purchased from LKT laboratories (St. Paul, MN). [¹²⁵I]iodoarylazidoprazosin (IAAP) was obtained from PerkinElmer Life and Analytical Sciences (Waltham, MA).

Cell Lines and Cell Culture.

The cell lines of the NCI Anticancer Drug Screen panel were obtained and grown in monolayers or in suspension in RPMI 1640 medium containing 10% fetal bovine serum, 2 mM glutamine, 100 units/ml penicillin, and 100 μg/ml streptomycin in 5% CO₂ at 37°. Human embryonic kidney (HEK) 293 cells (American Type Culture Collection, Manassas, VA) stably transfected with empty pcDNA 3.1 vector (Invitrogen, Carlstad, CA) (pcDNA) or vector containing full-length ABCG2 R-2 were maintained in Earle's minimal essential medium supplemented with 2 mg/ml G418 to enforce transporter expression (Robey et al., 2003). Validation studies also used ABCG2-overexpressing NCI-H460 MX20 cells that were obtained by step-wise selection and were maintained in 20 nM mitoxantrone (Robey et al., 2004). ABCG2-overexpressing MCF-7 FLV1000 cells were maintained in Richter's medium with 1000 nM flavopiridol (Robey et al., 2001). DNA isolated from the cell lines of the NCI drug screen was provided by the NCI DTP.

ABCG2 Functional Assay.

Functional assays with PhA were performed as described previously with minor modifications (Robey et al., 2004). In brief, trypsinized cells were incubated for 30 min at 37°C in 1 μM PhA in the presence or absence of 10 μM FTC, an ABCG2 inhibitor. The cells were subsequently washed and incubated at 37°C for 1 h in PhA-free medium continuing with FTC to block ABCG2-mediated efflux of PhA and generate the FTC/efflux histogram (Fig. 1A, dashed line) or continuing without FTC during the 1-h efflux period, generating the efflux histogram (solid line). PhA fluorescence was measured on a FACSort flow cytometer equipped with a 635-nm red diode laser. The difference in mean channel number between the FTC/efflux and efflux histogram, termed the inhibitable efflux, was calculated, and each cell line was tested at least twice. This value has been previously shown to correlate with ABCG2 expression (Robey et al., 2004). When the inhibitable efflux value was negative, it was assigned the value of zero.

This same method was used to test whether compounds identified in the NCI-ADS as potentially interacting with ABCG2 could inhibit transporter function. Potentially interacting compounds were incubated with wild-type ABCG2 transfected cells at a concentration of 10 μM to determine whether they inhibited PhA efflux. -Fold increase in PhA fluorescence was obtained by dividing the intracellular PhA fluorescence in the presence of each compound by the fluorescence in the absence of compound. Each compound was tested at least twice.

ABCG2 Expression Data.

Measures of ABCG2 gene expression were obtained from the NCI DTP. These results, obtained by polymerase chain reaction (Szakács et al., 2004) and gene expression microarrays (Lee et al., 2003), are publicly available (http://dtp.nci.nih.gov/mtargets/mt_index.html). Seven measures of ABCG2 mRNA were available and included in this study (NCI DTP identification numbers MT2678, GC14733, GC36729, GC56458, GC93477, GC152721, and GC228107).

ABCG2 SNP Genotyping.

The NCI 60 cell lines were genotyped for variations in the ABCG2 gene using DNA provided by the NCI DTP. After double stranded DNA content for each cell line was determined using the Quant-iT Picogreen dsDNA assay kit (Invitrogen, Carlsbad, CA) in conjunction with a FLUOstar Optima (BMG Labtech, Durham, NC) fluorescence plate reader, the identification of SNPs was performed using two different methods. First, genotype identification for seven variant sites in the ABCG2 gene was done for each cell line using the Affymetrix DMET platform as described previously (Dumaual et al., 2007). Second, genotyping for variations in intron 1 (rs2622604) was performed as follows: 20 ng of genomic DNA was concentrated, applied, and dried to a 384-well thermoplate (Thermo Fisher Scientific, Waltham, MA). Validated SNP identification primers for this variant along with VIC and 5-carboxyfluorescein probes and master mix for the TaqMan SNP Genotyping Assay (Applied Biosystems, Foster City, CA) were added to the wells, then sealed and processed on a 7900HT Fast Real-Time PCR System (Applied Biosystems) for 40 cycles (95°C for 15 s followed by 60°C for 60 s). SNP classification was performed with the onboard SDS 2.1 software and checked manually to ensure accuracy. The experiment was run in duplicate to confirm results.

National Cancer Institute Drug Screen Database and COMPARE Analysis.

The NCI Drug Screen database contains information on more than 140,000 compounds characterized for cytotoxicity patterns in the 60 human cancer cell lines. Cytotoxicity curves are generated, from which the GI₅₀ can be determined. The GI₅₀ is the time 0-corrected IC₅₀ value and is defined as the concentration of an agent that causes 50% growth inhibition. The GI₅₀ is different in each cell line, and the difference of that value from the mean GI₅₀ for the 60 cell lines for any given compound defines a cytotoxicity pattern that can be displayed as a “mean graph” or “fingerprint” (See Fig. 4). A vertical line represents the mean response of the cell lines to the test agent. The COMPARE database stores the screening data as the −log(GI₅₀) for historical reasons related to the mean graph sign conventions (Lee et al., 1994). This means that data with higher IC₅₀ values from more drug-resistant cell lines are stored with smaller (more negative) values and are found graphically to the left of the mean. Drug-sensitive cell lines are graphed to the right. A profile or seed probe can be entered to query the database for compounds that have a high positive or negative correlation. We generated a fingerprint of ABCG2 function in the 60 cell lines of the drug screen and used this as a seed to obtain compounds that have cytotoxicity profiles correlating with the ABCG2 function profile. These compounds were then subjected to further testing.

Competition of [¹²⁵I]IAAP Labeling.

Crude membranes (1 mg of protein/ml) from the MCF-7 FLV1000 cells were incubated in 50 mM Tris-HCl, pH 7.5, containing 20 μM test compound or FTC for 10 min at room temperature. Subsequently, 3 to 6 nM [¹²⁵I]IAAP (2200 Ci/mmol) was added, and the samples were incubated for an additional 5 min under subdued light. The samples were illuminated with a UV lamp (365 nm) for 10 min at room temperature (21–23°C). Labeled ABCG2 was immunoprecipitated by adding 800 μl of radioimmunoprecipitation assay buffer with 1% aprotinin followed by the addition of 10 μg of BXP-21 antibody (Kamiya Biomedical, Seattle, WA), after which the samples were incubated for 3 h at 4°C. The beads were pelleted by centrifuging at 13,000 rpm for 5 min at 4°C and then washed with radioimmunoprecipitation assay buffer in 1% aprotinin. SDS-PAGE sample buffer (25 μl) was then added, and the samples were incubated for 1 h at 37°C, followed by the addition of 25 μl of water and an additional incubation at 37°C for 30 min. Samples were separated by PAGE on a 7% Tris-acetate gel at constant voltage. The gel was dried and exposed to Bio-Max MR film (Carestream Health, Rochester, NY) for 3 to 6 days at −80°C. The incorporation of [¹²⁵I]IAAP into the ABCG2 band was quantified by estimating the radioactivity of this band using the STORM 860 PhosphorImager system (GE Healthcare, Chalfont St. Giles, Buckinghamshire, UK) and ImageQuaNT software (Shukla et al., 2006).

Cytotoxicity Assays.

Four-day cytotoxicity assays with sulforhodamine B were performed as described previously (Skehan et al., 1990). The cells were plated in flat-bottomed 96-well plates (10⁴ cells per well for transfected HEK 293 cells and 5 × 10³ cells per well for NCI-H460 parental and MX20-resistant cells) and allowed to attach for 24 h at 37°C. Compounds at various concentrations were added to the cells and the plates were allowed to incubate for 96 h at 37°C. The cells were subsequently fixed in 50% trichloroacetic acid and stained with sulforhodamine B solution (0.4% sulforhodamine B in 1% acetic acid). Optical densities were read on a plate reader (Bio-Rad Laboratories, Hercules, CA) at an absorbance of 540 nm. Each concentration was tested in quadruplicate. Combination studies with putative potential ABCG2 inhibitors and topotecan, a known ABCG2 substrate, were performed in wild-type ABCG2-transfected and empty vector-transfected cells using this same method. Relative resistance values were calculated by dividing the IC₅₀ of the ABCG2-expressing line by its corresponding parental line.

Results

ABCG2 Function and the NCI-60 Cell Lines.

Relative ABCG2 transporter function in 59 of the 60 cell lines of the drug screen was measured by flow cytometry using the PhA assay. Cells were incubated with PhA in the absence of FTC to generate the efflux histogram (solid line), whereas cells were incubated in PhA in the presence of FTC to generate the FTC/efflux histogram (dashed line) as outlined under Materials and Methods (Fig. 1A). The difference between the two histograms has been shown to correlate with levels of cell surface ABCG2 expression (Robey et al., 2004). Figure 1A shows results from two cell lines found to have the highest levels of ABCG2 function among cells in the screen (A549 and NCI-H460), one cell line with moderate function (NCI-H23), and one with nearly undetectable ABCG2 function (OVCAR4).

Figure 1B summarizes the results from 59 of the 60 cell lines, expressed as the difference in mean channel number between the FTC/efflux and the efflux histograms. This difference reflects the degree of ABCG2 transporter-mediated efflux and thus is a measure of relative ABCG2 function in each cell line. Although levels of ABCG2-mediated PhA efflux were relatively low compared with levels observed in drug-selected cell lines, the highest levels of ABCG2 function were found in the NCI-H460, A549, RPMI-8226, NCI-H23, HCC2998, KM12, and SF295 cell lines. Lower levels were found in IGROV1 and LOX-IMVI, and ABCG2 function was nearly undetectable in HL-60, OVCAR4, and SNB-19 cells.

The NCI-60 cell lines represent nine different tumor types, including leukemia, non–small-cell lung cancer, colon cancer, central nervous system tumors, melanoma, ovarian carcinoma, renal cancer, prostate cancer, and breast cancer. Relative ABCG2 function by tissue type within the DTP cell lines is shown in Fig. 1C. With the exception of four cell lines among the leukemia (RPMI-8226) and lung cancer subsets (A549, NCI-H460, and NCI-H23), the range of ABCG2 function was low across all tumor types.

Correlation between Expression and Function.

Various measures of ABCG2 mRNA available from the NCI Molecular Targets Database were used to determine a correlation between expression and ABCG2 function (MT2888). ABCG2 mRNA expression was analyzed by reverse transcription-polymerase chain reaction (MT2678) as well as in gene expression arrays, including the Affymetrix U95 (GC36729, GC56458, and GC93477), the Affymetrix U133 (GC152721 and GC228107), and a customized oligonucleotide expression array (GC14733). We constructed a correlation matrix to determine the relationships between ABCG2 function, ABCG2 mRNA expression, and ABCG2 expression determined by the gene expression arrays (Supplemental Table 1). Pearson correlation coefficients between functional data and expression data ranged from 0.42 (GC14733) to 0.89 (GC228107), suggesting consensus between arrays and the functional data [except for the Affymetrix U95 A chip (GC94377)]. Consensus with the latter and any of the other variables was low (correlation coefficient of −0.05 to 0.14), but data from this chip are generally considered unreliable.

SNPs in the NCI-60 Cell Lines.

Because SNP variants are known to impair ABCG2 expression and function, cells expressing these variants may confound results based on mRNA expression. Thus, we evaluated ABCG2 genotypes in the cell lines of the drug screen. DNA from 59 of the NCI-60 cell lines were genotyped to identify the presence of the Q141K allelic variant, previously shown to have diminished function compared with wild-type ABCG2 (Imai et al., 2002; Morisaki et al., 2005), as well as a SNP variant in the first intron of the gene recently found to correlate with chemotherapy toxicity in vivo (Rudin et al., 2008). An additional six variant sites were also genotyped (Table 1). For the Q141K SNP, 12 cell lines were found to contain a variant allele. Two lines were homozygous for this variant (LOX IMVI and A498). The other 10 lines were heterozygous, containing one Q141K variant allele (A549, COLO205, HCT116, SF295, MALME-3M, SK-OV-3, CAKI-1, HOP62, HOP92, and MDA-MB-231).

Twenty cell lines contained the intron one variant. Four cell lines were homozygous variant (SW620, OVCAR 5, BT549, and T47D), whereas 16 were heterozygous variant (MOLT4, HOP-92, HCC-2998, SF539, SNB19, SNB75, U251, SKMEL5, OVCAR3, OVCAR8, 786-0, RXF393, TK10, NCI ADR-RES, MDA-MB-231, and HS578T). For the other six variant sites, all 60 cell lines were homozygous wild type at each SNP site.

To determine whether these variants affected the correlation between gene expression and transporter function, the scatter plot analysis comparing expression with function was repeated after excluding those cell lines with the Q141K SNP or the intron 1 SNP. There was no clear pattern of improvement in correlation across these measures of gene expression and ABCG2 function (data not shown).

COMPARE Analysis.

We next used the COMPARE program to probe the drug sensitivity database using the ABCG2 function profile. Because higher ABCG2 expression or function should lead to higher cellular resistance to compounds effluxed by the drug transporter, we hypothesized that our profile of ABCG2 levels across the 60 cell lines could be correlated with the cytotoxicity profiles of substrate drugs, as was observed for P-glycoprotein (Lee et al., 1997; Alvarez et al., 1998). The COMPARE correlation analysis produces a Pearson correlation coefficient (PCC) correlating ABCG2 levels with the sensitivities of the 60 cell lines to the compounds in the database.

In the COMPARE program, search parameters can be adjusted, including the requirement for a minimum number of cell lines that correlate with the probe and a minimum S.D. Through repeated testing and comparing the fingerprints of identified agents, we determined that optimized settings were: 30 cell lines and S.D. > 0.1. Because the majority of cell lines had modest and similar functional measures, we chose the lower setting of 30 cell lines (of 60) with which the ABCG2 fingerprint and a compound had to correlate to give a positive result. The S.D. of > 0.1 ensured that compounds had a significant variation across the 60 cell lines, and that a high PCC value showing correlation between the compound and ABCG2 function was not driven by a single or a few cell lines.

We probed the drug sensitivity database using the profile produced by the ABCG2 functional data to identify compounds with a high positive or negative correlation. The PCC is highly positively correlated with values approaching 1.0 and highly negatively correlated with values approaching −1.0. Values that range between −0.3 and +0.3 are thought to be not significantly correlated in the screen. Figure 2, right, shows the ABCG2 fingerprint generated by measuring FTC-inhibitable efflux in the cell lines. This fingerprint was used as the seed in the COMPARE algorithm.

From the NCI DTP web-accessible database, we obtained the cytoxicity profiles of 175 commonly used chemotherapy agents whose mechanisms are well understood, included as the NCI Anticancer Drug Screen's Standard Agents Database. This profile includes the GI₅₀ of a compound for each of the 60 cell lines. Using COMPARE, we related the cytotoxicity profile for each standard agent to ABCG2 function in each cell line using the Pearson's correlation coefficient as well as the Spearman's rank correlation coefficient. There was no significant correlation between the ABCG2 functional fingerprint and the cytotoxicity profiles. As seen in Fig. 3, compounds known to be ABCG2 substrates such as mitoxantrone and topotecan had surprisingly lower than expected correlations of 0.082 and 0.011, respectively. The few exceptions to the generally low PCC values were a significant positive correlation with the agent pancratistatin (PCC = 0.447) and a negative correlation with ftorafur (PCC = −0.437). Reanalysis omitting cell lines known to express high levels of P-gp or MRP1 did not improve PCCs (data not shown).

The ABCG2 functional fingerprint was then used as a probe in the larger collection of more than 100,000 synthetic agents that have been studied in the NCI 60. The compounds vary in terms of how well they are characterized, both molecularly and as potential anticancer agents. With this larger collection, only 187 compounds were found to have a positive correlation greater than 0.4 or a negative correlation less then −0.4. One compound (NSC651644) was found to have a PCC value above 0.6, four compounds had values above 0.5 (NSC722812, NSC686342, NSC625546, and NSC114609), and 66 compounds had correlations measuring between 0.4 and 0.5 (Table 2). A positive correlation in this setting could indicate that these compounds were ABCG2 substrates. The fingerprint of one compound with a positive PCC (0.456), NSC107392, is shown in Fig. 2, left.

Also found from this COMPARE analysis were compounds with negative correlations. Among the compounds found to have a high negative correlation with the ABCG2 functional fingerprint was NSC651424 (PCC = −0.739). Two additional compounds had correlations between −0.7 and −0.6, and another 23 compounds had PCC values between −0.6 and −0.5 (Table 2). Negative correlation in this context suggests that these agents are preferentially toxic to cells that express ABCG2 and could be considered potential ABCG2 “targeting” agents. This was found to be true in the case of P-gp, where some compounds with negative PCCs were found to be more toxic to cells overexpressing the transporter (Alvarez et al., 1995; Szakács et al., 2004).

Available compounds with high positive and low negative PCCs were obtained from the NCI DTP and subjected to further testing. One limitation at this step was that many compounds with high correlations were not available, including NSC651424 (PCC = −0.739), 693023 (−0.663), and 671546 (−0.626). A total of 27 compounds were obtained, 17 with positive PCCs and 10 with negative PCCs (Table 3). We selected available compounds with a PCC ≥ 0.4 or ≤ −0.4 because these compounds generally had p values <0.005. Compounds with PCCs ≤ 0.35 tended to have p values greater than 0.05. In the case of some compounds, the PCC values shown represent the average PCC from several experiments in the drug screen.

Characterization of 27 Compounds Obtained by COMPARE Analysis.

To determine whether any of the compounds were substrates of the ABCG2 transporter, 4-day cytotoxicity assays were performed with the 27 compounds on HEK293 cells transfected with empty pcDNA3.1 vector or vector containing DNA encoding the wild-type ABCG2 transporter. Relative resistance values were calculated by dividing the IC₅₀ for each compound in the ABCG2-transfected cells by the IC₅₀ for each compound in the empty vector-transfected cells. In cytotoxicity assays, ABCG2-expressing HEK cells are more resistant to a substrate than cells transfected with empty vector, resulting in a higher relative resistance value. ABCG2-expressing HEK cells were 19-fold resistant to topotecan compared with empty vector-transfected cells (Table 3). Substrates, identified as those compounds for which there was a greater than 2-fold relative resistance value, included 5 of the 27 compounds tested (NSC107392, NSC265473, NSC305458, NSC349156, and NSC382054) as listed in Table 3. ABCG2-transfected cells were particularly resistant to compounds NSC107392, NSC265473, and NSC349156, with relative resistance values greater than 37-fold, suggesting that these compounds are readily transported by ABCG2. During the course of our study, we discovered that NSC265473 and NSC305458 are actually the same compound, providing a serendipitous internal control.

We next sought to determine whether resistance to the test compounds could be conferred by the lower endogenous levels of ABCG2 found in the unselected drug screen cell lines. Four-day cytotoxicity assays were performed on A549, NCI-H460, and HT29 cells with SN-38, topotecan, NSC107392, or NSC265473 in the presence or absence of 5 μM FTC, an ABCG2 inhibitor. Dose-modifying factors, representing sensitization as a result of incubation in the presence of FTC, were then calculated by dividing IC₅₀ values obtained in the absence of FTC by IC₅₀ values obtained in the absence of FTC. In general, dose-modifying factor values were lower for SN-38 (range, 1–2.7) and topotecan (range, 1–2.6) than for the two novel substrates NSC107392 (range, 1.4–3.0) and NSC265473 (range, 2.1–6.5). However, these differences did not achieve statistical significance.

The negatively correlating compounds were expected to be more toxic to cells expressing ABCG2. However, only 2 of the 27 tested compounds (NSC103054 and NSC174939) showed evidence of greater cytotoxicity in ABCG2-expressing cells with relative resistance values less than 0.5. One of two compounds (NSC103054) that was more toxic to ABCG2-transfected cells had a positive PCC (PCC = 0.471). Four of the 10 compounds with negative PCCs were not toxic to either of the two transfected cell lines at concentrations of up to 100 μM.

We also performed cytotoxicity assays with a subset of the 27 compounds on parental NCI-H460 and ABCG2-overexpressing NCI-H460 MX20 cells. As seen in supplemental Table 2, similar relative resistance values were observed and did not differ from those for the transfected cells (Table 3) by more than 3-fold except in the case of NSC349156, where the transfected cells had relative resistance values that were more than 4-fold higher than for the NCI-H460 and NCI-H460 MX20 cells.

Next, we sought to determine whether these 27 compounds were able to interact with ABCG2 by testing their ability to inhibit photo-cross-linking of ABCG2 with [¹²⁵I]IAAP as described above. This method has been used to identify compounds that bind to ABCG2 at the same site as prazosin, a known ABCG2 substrate, by measuring their ability to prevent binding of the radiolabeled, photo-cross-linkable analog of prazosin, [¹²⁵I]IAAP, to ABCG2. An example of the gel results from this assay is shown in Fig. 4A. Results for each compound are listed in Table 3, with results reported as the percentage inhibition of ABCG2 labeling (i.e., 80% inhibited for FTC). Eight of the 17 compounds with a positive PCC inhibited IAAP binding by more than 50% (NSC651644, NSC103054, NSC623636, NSC620515, NSC691417, NSC620303, NSC297093, and NSC265473). In addition, 5 of 10 compounds with a negative PCC inhibited IAAP binding by more than 50% (NSC153330, NSC45384, NSC24048, NSC382054, and NSC608001).

As another measure of interaction with ABCG2, positively or negatively correlating agents were tested for their ability to inhibit ABCG2 function by determining the effect of 10 μM concentrations of the compounds on the efflux of PhA from wild-type ABCG2-transfected cells. ABCG2-transfected cells were incubated in PhA in the presence (dashed line) or absence (solid line) of 10 μM concentrations of each compound, and the -fold-increase in PhA fluorescence in the dashed histogram compared with the solid histogram was calculated. Results obtained with five of the tested compounds and the positive control FTC are given in Fig. 4B; results for all compounds are summarized in Table 3. Three of 10 compounds tested with a negative PCC were found to increase PhA fluorescence by 2-fold or more (NSC24048, NSC45384, and NSC608001). In addition, three of 17 compounds with a positive correlation were found to increase intracellular PhA fluorescence by 2-fold or greater (NSC636795, NSC103054, and NSC691417). Structures for selected test compounds are provided in Fig. 4C. Structures of other compounds can be obtained from http://dtp.nci.nih.gov.

Of the six compounds that were able to increase PhA fluorescence in the ABCG2-transfected cells by at least 2-fold, four compounds that were nontoxic at concentrations shown to inhibit ABCG2 function were further tested in combination studies with topotecan, a cytotoxic agent known to be a ABCG2 substrate. Cytotoxicity asssays were performed on ABCG2-transfected and empty vector-transfected HEK293 cells with topotecan at various concentrations, along with nontoxic concentrations of the compounds NSC608001 (10 μM, PCC = −0.409), NSC636795 (10 μM, PCC = 0.477), NSC153330 (25 μM, PCC = −0.502), and NSC24048 (10 μM, PCC = −0.472). In all cases, the concentrations chosen resulted in less than 10% cell death of both the empty vector-transfected and the ABCG2-transfected cells. Although NSC608001, NSC636795, and NSC153330 partially reversed ABCG2-mediated topotecan resistance in ABCG2-transfected cells, NSC24048 completely reversed ABCG2-mediated resistance (Table 4).

Discussion

The NCI-ADS has proven to be a powerful basic and translational research tool to identify new anticancer agents, characterize cytotoxic drug activity, and investigate mechanisms of cancer drug resistance. We and others have previously used this research tool and database to identify agents that targeted the epidermal growth factor receptor (Wosikowski et al., 1997; Liu et al., 2007), the erbB2 pathway (Wosikowski et al., 1997) and Chk2 (Jobson et al., 2007), as well as agents that were substrates of the drug transporters P-gp (Lee et al., 1994) and MRP1 (Alvarez et al., 1998).

In the current study, we characterized the function of the ABCG2 efflux transporter across the 60 cell lines of the NCI-ADS. Our hypothesis was that a measure of transporter function might be more sensitive than gene expression to identify compounds that interact with ABCG2, a strategy that in the past has been used successfully to identify substrates for other ABC transporters. Using ABCG2 gene expression measured by real-time reverse transcription-polymerase chain reaction as a fingerprint probe, Szakács et al. (2004) did not identify transporter substrates within the drug screen. As shown here, using transporter function as a fingerprint to probe the screen did identify novel substrates and inhibitors of this critical drug transporter.

Although we did identify potential substrates and targeting agents using transporter function as the seed probe entered into the COMPARE program, in general we were unable to identify as many compounds as were identified previously when P-glycoprotein function was used as the seed (Lee et al., 1994). Using PhA efflux as a measure of ABCG2 function, only one compound was found to have a PCC value above 0.6, only four compounds had values above 0.5, and 66 compounds measured between 0.4 and 0.5. Although these values are low compared with correlations obtained using P-gp function as the seed, where PCC values for the top 20 compounds ranged from 0.816 to 0.976, they are in the same range as that observed for several molecular targets, including MRP1 (Alvarez et al., 1998) and epidermal growth factor receptor (Wosikowski et al., 1997). Known substrates of ABCG2 that are found in the drug screen, such as mitoxantrone, topotecan, and SN-38 (the active metabolite of irinotecan), were surprisingly not found to significantly correlate with ABCG2 function, leading us to the conclusion that the screen cannot be used as a predictive tool to determine whether or not a compound is a substrate. We were, however, able to identify the novel ABCG2 substrates NSC107392, NSC265473, and NSC349156 (pancratistatin). Pancratistatin is under investigation for its antitumor properties (McLachlan et al., 2005). It is noteworthy that a previously reported P-gp targeting agent, a thiosemicarbazone, was also shown to be an ABCG2 substrate (Wu et al., 2007).

There are several possible explanations for the somewhat low PCC values obtained and why known substrates were not identified. First, the range of ABCG2 function across the cell lines in the screen is not as broad as that for P-gp. When values are widely separated across the 60 cell lines, stronger correlations are possible. Few cell lines have relatively high function of ABCG2 as measured by PhA efflux (A549, NCI-H460, RPMI 8226, and NCI-H23). This compares to 12 cell lines found to have high P-gp function in our previous work, some at levels found in drug-selected cell lines (Lee et al., 1994). The generally low levels of ABCG2 expression in the cell lines of the drug screen would also serve to reduce correlations for compounds that are relatively poor substrates of ABCG2 such as flavopiridol, because high levels of ABCG2 expression are needed to confer resistance to such compounds. Ultimately, including more cell lines with higher levels of ABCG2 in the drug screen might allow for identification of known ABCG2 substrates. Second, the contribution of the drug target might obscure the contribution of ABCG2 to resistance. Cancer cells that express high levels of a drug target, such as topoisomerase, may be exquisitely sensitive to drugs targeting the protein such that ABCG2 is not able to keep intracellular drug concentrations low enough to confer resistance. Likewise, ABCG2 substrates such as imatinib would not be identified as substrates by the drug screen unless cell lines expressed the BCR-ABL fusion protein or an activating KIT mutation. Third, expression of other transporters, such as P-gp or MRP1, may obscure the contribution of ABCG2 to drug resistance in a given cell line. Some in vitro studies support such a hypothesis, because, for example, topotecan uptake into the brain was not observed to be increased in Abcg2-deficient mice until the Mdr1/2 genes were also deleted (de Vries et al., 2007). Another possibility is that SNPs may directly affect ABCG2 function. A polymorphism that impairs function could readily explain the failure of gene expression data to correlate with cytotoxicity data. However, a functional measure should account for any influence of polymorphic forms of ABCG2 as we presently understand them. Furthermore, our analyses on the correlation between function and expression, as well as the correlation between compound cytotoxicity and function, did not improve when genetic variant cell lines were excluded. Finally, ABCG2 localization may have an impact on drug resistance. Intracellular expression of ABCG2 has been observed in some tissue samples, and it is unknown whether intracellular ABCG2 can mediate drug resistance. If so, measuring PhA efflux might underestimate the contribution of ABCG2 to drug resistance in some cell lines. If this were true, measurement of ABCG2 by immunoblot should improve correlations; again, however, levels of ABCG2 remain generally low in the 60 cell lines of the screen. Nonetheless, measurement of ABCG2 function did allow us to identify novel compounds that interact with ABCG2.

In this analysis we also looked for correlations with PCC values that were negative to identify compounds that might target cells that express high levels of ABCG2. Of the 27 compounds that we examined, only two were slightly more toxic (2.5- and 3-fold) in ABCG2-transfected cells versus empty vector transfected cells. Because the COMPARE program was able to identify compounds with negative PCCs that selectively target cells that overexpress P-gp (Szakács et al., 2004), it is possible that similar compounds that are selectively toxic to ABCG2-overexpressing cells could be identified.

ABCG2-targeting compounds may be of interest for a novel clinical approach—that of targeting putative cancer stem cells. Normal hematopoietic stem cells as well as cancer stem cells are identified by flow cytometry in a distinct population of cells that stain dimly with the fluorescent dye Hoechst 33342. This so-called “side population” is due to ABCG2-mediated Hoechst efflux and is enriched in tumorigenic or potential cancer stem cells (Zhou et al., 2001). Whether true cancer stem cells exist or not is still hotly debated, but it does appear that cells in this side population could have intrinsic chemotherapy resistance as a result of ABCG2-mediated drug efflux. Therefore, a potential therapeutic modality would be to use compounds that are selectively toxic to this subset of cells.

Of the 27 compounds tested, six were found to inhibit ABCG2-mediated PhA transport in ABCG2-transfected cells at 10 μM. In particular, NSC24048, NSC153330, NSC608001, and NSC636795 were sufficiently nontoxic that combination cytotoxicity assays could be performed. It is noteworthy that these included compounds with both positive and negative PCCs. NSC24048 was the most potent compound tested and completely reversed ABCG2-mediated topotecan resistance in the ABCG2-transfected cells at a concentration of 10 μM. The identification of potential inhibitors of the ABCG2 drug transporter may prove to be clinically useful given the central role that ABCG2 plays in the blood-brain barrier and in mediating oral absorption of drugs. ABCG2 inhibitors may provide increased delivery of chemotherapy agents into sanctuary sites such as the brain to treat primary or metastatic disease (Breedveld et al., 2005).

In conclusion, measurements of ABCG2 transporter function were more successful than gene expression when used as a probe to identify drugs that interact with this transporter. We identified novel cytotoxic agents that are substrates of ABCG2, as well as some drugs that functioned as transport inhibitors. Further investigations and preclinical evaluation of these substrates and inhibitors are warranted and are ongoing.