Abstract
Since the early 1990s the Developmental Therapeutics Program of the National Cancer Institute (NCI) has utilized a panel of 60 human tumor cell lines (NCI60) representing 9 tissue types to screen for potential new anticancer agents. To date, about 100,000 compounds and 50,000 natural product extracts have been screened. Early in this program it was discovered that the pattern of growth inhibition in these cell lines was similar for compounds of similar mechanism. The development of the COMPARE algorithm provided a means by which investigators, starting with a compound of interest, could identify other compounds whose pattern of growth inhibition was similar. With extensive molecular characterization of these cell lines, COMPARE and other user-defined algorithms have been used to link patterns of molecular expression and drug sensitivity. We describe here the results of screening current Food and Drug Administration (FDA)-approved anticancer agents in the NCI60 screen, with an emphasis on those agents that target signal transduction. We analyzed results from agents with mechanisms of action presumed to be similar; we also carried out a hierarchical clustering of all of these agents. The addition of data from recently approved anticancer agents will increase the utility of the NCI60 databases to the cancer research community. These data are freely accessible to the public on the DTP website (http://dtp.cancer.gov/). The FDA-approved anticancer agents are themselves available from the NCI as a plated set of compounds for research use. Mol Cancer Ther; 9(5); 1451–60. ©2010 AACR.
Introduction
The Developmental Therapeutics Program (DTP) of the National Cancer Institute (NCI) has as its mission the discovery and development of novel anticancer agents. The program started more than 50 years ago as the Cancer Chemotherapy National Service Center and has had a significant role in the development of many agents that are now part of standard cancer care. Notable examples include paclitaxel (Taxol; ref. 1) and bortezomib (Velcade; ref. 2). One of the tools that the DTP employs in the early stage of drug discovery and development is the NCI60 cell line screen, which utilizes a panel of 60 human tumor cell lines, chosen for their ability both to perform consistently under the conditions of the assay and to represent a variety of tumor types (3). With the NCI60 in use since the early 1990s, nearly 100,000 compounds and 50,000 natural product extracts have been examined for therapeutic activity in this assay. About half of the compounds are covered by confidentiality agreements. Data for the remaining compounds are freely available through the DTP website (4) for independent analysis by any investigator.
In addition to the large body of compound sensitivity data, this panel of cell lines has been extensively characterized at the molecular level by numerous groups throughout the world (5). The resulting data are publicly available through the DTP website (6). RNA expression analysis, derived from six microarray measurements on four different platforms, provides data on >100,000 features for each cell line (7). Karyotype analysis of the cells has revealed numerous alterations in chromosome number and organization (8). Single nucleotide polymorphisms were determined on high-density arrays, which also provide estimates of DNA copy number for 120,000 sites (9). Additional molecular characterization includes microRNA expression (10, 11), DNA mutation (12), protein analysis (13), DNA methylation (14), functional target analysis (15), and metabolomic analysis.
Both the compound sensitivity database and the molecular characterization data provide a rich context for the interpretation of novel compounds and targets in the NCI60. The COMPARE algorithm (16) is provided on the website to enable investigators to search for compounds or molecular targets with similar patterns of sensitivity or expression in the cell line screen. Because the data can be readily downloaded from the website, researchers may apply their own algorithms for data analysis. The addition of these Food and Drug Administration (FDA)-approved anticancer agents to the dataset increases the utility of these databases to the cancer research community.
Materials and Methods
Compounds
All compounds were obtained from the NCI DTP Repository (Rockville, MD). Plated sets of approved oncology drugs are also available to researchers on request via the DTP website (17). All proprietary agents were purchased commercially. When necessary, the active ingredient was extracted from formulated material, and purified. All compounds were assayed to confirm potency and purity, and these data are available at the website above.
NCI60 anticancer drug screen
The screening methodology has been described in detail elsewhere (18). Briefly, cells are seeded in 96-well plates at an appropriate density and are incubated for 1 day, after which some of the plates are processed to determine a time zero density. To the remaining plates, compounds are added over a 5-log mol/L concentration range. Plates are incubated a further 2 days, then fixed and stained with sulphorhodamine B. Growth inhibition is calculated relative to cells without drug treatment and the time zero control. The use of a time zero control allows the determination of cell kill as well as net growth inhibition.
If a particular end point falls outside of the testing range for a given cell line, the database assigns a value equal to either the highest or the lowest concentration tested. For a potent compound, such that growth inhibition in a given cell line is >50% at all concentrations, the GI50, which is the concentration of a compound that causes 50% growth inhibition relative to the no-drug control, would be imputed as the lowest concentration tested. For a relatively inactive compound, such that a given cell line is inhibited <50% at all concentrations, the GI50 is assigned as the highest concentration tested.
The cell lines used in the screen have been extensively molecularly characterized, including high-density single nucleotide polymorphism genotyping. More recently genotyping has been done with the AmpFlSTR Identifiler PCR Amplification kit (Applied Biosystems), with results consistent with published results from others (19).
Statistical analyses
Analyses were done using JMP7 statistical software (SAS Institute Inc.). Pairwise Pearson correlation coefficients (PCC) were calculated using the multivariate platform. Hierarchical clustering was done using the Ward method.
Results
The majority of the FDA-approved anticancer drugs were tested in the NCI60 screen at least twice. If the initial assay was done over a nonoptimal concentration range, additional assays were done in concentration ranges that better captured the performance of the compound. From the dose-response curves, three end points were calculated, as illustrated in Fig. 1, using dose-response data for dasatinib in the melanoma cell line panel: the GI50; the TGI (total growth inhibition), which is the concentration that yields no net growth over the course of the assay; and the LC50, which is the concentration that kills 50% of the cells that were present at the time of drug addition.
Dose-response graphs for dasatinib assayed in the melanoma panel, showing end point calculations. Dasatinib (NSC 732517) was tested at five concentrations (1 log dilutions from 10−4 mol/L to 10−8 mol/L). Growth percent of 100 corresponds to growth seen in untreated cells. Growth percent of 0 indicates no net growth over the course of the assay (i.e., equal to the number of cells at time zero). Growth percent of -100 results when all cells are killed. Three end points are routinely calculated: (a) GI50, the log mol/L concentration yielding a growth percent of 50 (i.e., 50% growth inhibition); (b) TGI, or total growth inhibition, the log mol/L concentration yielding a growth percent of 0; and (c) LC50, the log mol/L concentration yielding a growth percent of -50, or lethality in 50% of the starting cells. These end points are illustrated for cell line LOX-IMVI (red open circle). Other cell lines displayed are Malme-3M (red open diamond), M14 (red open triangle), MDA-MB-435 (red open square), SK-MEL-2 (solid blue circle), SK-MEL-28 (solid blue diamond), SK-MEL-5 (solid blue triangle), UACC-257 (solid blue square), and UACC-62 (open green circle).
Table 1 presents the mean sensitivity for each compound across all cell lines. When multiple concentration ranges were examined, the data were manually inspected to determine the optimal range to use. This process was done separately for each of the three end points, because the optimal concentration range for determining GI50 may differ from that for determining LC50. If many cell lines were outside the testing range for a given end point, the value in Table 1 was estimated as > or < than the extremes of the concentration range.
US Food and Drug Administration–approved anticancer agents: activity in the NCI60 panel
The GI50 values for each approved drug were used to calculate Pearson correlation coefficients between all of the other drugs. When multiple experiments were available for a drug in a given concentration range, the values for that drug were averaged to obtain a mean for each cell line. If multiple concentration ranges were tested, the dose response curves were visually inspected to determine which range provided the most reliable GI50 values, and those were used for the correlation analysis. If multiple concentration ranges seemed to be acceptable, all were used separately for the correlations. The resulting “matrix COMPARE” correlations were then hierarchically clustered; the results are shown graphically in Fig. 2. Drugs were color coded according to mechanism of action. Most agents cluster with other drugs of similar presumed mechanism.
Clustering of correlations of NCI60 GI50 patterns for all drugs. Pearson correlation coefficients comparing the GI50 patterns of each drug with all other drugs were hierarchically clustered. The agents are color-coded according to mechanistic category: Purple, signaling agents; blue, alkylating and other DNA damaging agents; turquoise, tubulin binders; orange, topoisomerase poisons; green, antimetabolites and nucleosides; red, hormonal agents; gray, all others. The correlation underlying this clustering can be found in Supplementary Table S1, presented in the same sort order as this figure.
Eleven drugs that affect signal transduction were tested in the NCI60 screen. The mean potency of these compounds was quite variable. The most potent were bortezomib (mean GI50 of 0.51 nmol/L) and temsirolimus (mean GI50 of 38 nmol/L). The mean potency of the kinase inhibitors ranged from 0.3 mmol/L (dasatinib) to 15 mmol/L (imatinib). However, because of their high target specificity, kinase inhibitors may inhibit the growth of only a small number of cell lines. Imatinib is the most extreme example, as shown in Fig. 3. Although the mean GI50 is 15 mmol/L, the only cell line in the panel bearing the BCR-ABL translocation (K-562) is roughly 1,000-fold more sensitive (GI50 of 0.02 mmol/L).
Dose response graphs for all cell lines in the NCI60 panel exposed to imatinib (NSC 743414). Imatinib was tested at five concentrations (1 log dilutions from 10−4 mol/L to 10−8 mol/L). Note that only one of the cell lines, K-562, which harbors a BCR-Abl gene fusion, has significant sensitivity to this BCR-Abl/KIT/PDGFR inhibitor. The GI50 and TGI concentrations for K-562 are indicated. Imatinib did not cause sufficient lethality in this cell line to calculate LC50. The graph is color-coded by tissue of origin: red, leukemia cell line; blue, lung cancer; green, colon cancer; gray, central nervous system cancer; coral, melanoma; purple, ovarian cancer; gold, renal cancer; turquoise, prostate cancer; pink, breast cancer.
Bortezomib was tested in the NCI60 screen during its early development, along with other related antiproteasomal drug candidates. These agents had a unique signature in the NCI60 screen – they were what may be referred to as “COMPARE negative,” that is, their pattern of growth inhibition did not resemble other previously tested classes of compounds. In addition, the potency of the series of proteasome inhibitory compounds in the NCI60 panel was proportional to activity against purifiedproteasomes (20). One of the more sensitive cell lines in the panel was the multiple myeloma line RPMI-8226; bortezomib is now a standard of care for the treatment of myeloma. Figure 4 shows the NCI60 data for bortezomib in two different formats: as a waterfall plot of all TGI data (Fig. 4A), and as a dose-response plot of all cell lines (Fig. 4B). These graphs show that bortezomib has a particular growth inhibition signature, with some cell lines being exquisitely sensitive, others being relatively resistant.
NCI60 graphs for bortezomib (NSC 681239). The data for bortezomib tested at five concentrations (1 log dilutions from 10−6 mol/L to 10−10 mol/L) are presented in two different formats. A, “waterfall” plot of GI50 molar values, with the most sensitive cell lines for each end point at the top of the graph. B, dose-response curves for all cell lines overlaid on the same plot. Cell lines are color-coded as for Fig. 3.
Dasatinib also showed a fairly unique pattern of NCI60 activity (Fig. 1). Interestingly, this pattern did not have a significant correlation (PCC = 0.15) with that of imatinib, which targets BCR-ABL, c-KIT, and platelet-derived growth factor receptor (PDGFR). Dasatinib inhibits these as well as Src family kinases. Imatinib is highly specific for the BCR-ABL cell line K-562, as shown in Fig. 3; none of the NCI60 cell lines harbors a KIT mutation. Although dasatinib is also highly active against K-562, it is also active against many other cell lines in the panel, as shown in Fig. 1. Among the most sensitive cell lines were those expressing higher levels of both PDGFRA and PDGFRB (data not shown). The central nervous system line U251, which expresses PDGFRA, but not PDGFRB, was relatively insensitive to dasatinib.
A matrix COMPARE analysis was done for the signaling agents, as described above, and the resulting Pearson correlation coefficients were clustered. The results are shown in Fig. 5. All three drugs that target epidermal growth factor receptor (EGFR) cluster together, including lapatinib, an agent that also targets ERBB2.
Clustering of correlations of NCI60 GI50 patterns for the signaling drugs. PCCs for the agents targeting signal transduction were hierarchically clustered in two symmetric dimensions. A heat map of the PCCs is shown, with higher correlations in red and lower PCCs in blue.
Mean graphs are shown in Fig. 6 for gefitinib and lapatinib. The graphs are visually similar, and COMPARE analysis confirms this, with a high correlation (PCC of 0.88) between these agents. Several cell lines are particularly sensitive to these two agents (lung EKVX, lung NCI-H322M, ovarian IGROV1, ovarian SK-OV-3, renal ACHN, renal TK-10, breast MDA-MB-468); all of these lines are KRAS wild-type (21), in agreement with clinical findings that KRAS mutant tumors are unresponsive to these drugs (22).
Mean graph plots of GI50 values for gefitinib (NSC 715055) and lapatinib (NSC 745750). GI50 values for each cell line were calculated from dose-response curves. The mean GI50 for each compound across all 60 cell lines was calculated. The difference between the GI50 for a particular cell line and the mean GI50 is plotted. Cell lines that were more sensitive are displayed as bars that project to the right of the mean. Cell lines that were less sensitive are displayed with bars projected to the left. Cell lines are color-coded as in Fig. 3. Mean graphs for two compounds with similar mechanisms are shown. Both gefitinib and lapatinib inhibit the tyrosine kinase EGFR, and lapatinib also inhibits the related kinase ERBB2. The two compounds give similar mean graph patterns. The degree of similarity was quantitated using the COMPARE algorithm, which gave a PCC of 0.88, confirming that these patterns are very similar. The most responsive cell lines to these agents are all wild type for KRAS, in line with what has been observed in the clinic.
Imatinib and nilotinib, which inhibit the BCR-ABL kinase, as well as KIT and PDGFR, cluster with one another, whereas dasatinib, which targets Src family kinases as well, does not cluster with these two BCR-ABL inhibitors. The mTOR inhibitors temsirolimus and everolimus cluster with one another, and are distinct from the other signaling agents.
Epigenetic regulation of gene expression has been implicated in the regulation of many cancer-related genes. This has inspired the development of multiple agents that inhibit histone deacetylases (HDAC), and two are currently approved oncology drugs. Vorinostat is a broad HDAC inhibitor that blocks the action of class I, II, and IV HDACs, whereas romidepsin has a narrower profile, inhibiting class I HDACs (reviewed in ref. 23). The GI50 patterns for vorinostat and romidepsin are not similar.
Previous publications have reviewed the NCI60 results from traditional cytotoxics, such as tubulin-directed agents, alkylating agents, and topoisomerase poisons. These data have been used to identify new compounds with similar patterns of growth inhibition to agents of known mechanism. Starting with the topoisomerase I poison camptothecin, Kohlhagen et al. utilized COMPARE to identify a novel structural class, the indenoisoquinolines, as topoisomerase poisons (24). After testing many analogs, two indenoisoquinolines have started clinical trials (25).
Two recently approved anticancer agents can be classified as traditional cytotoxics. Ixabepilone, a tubulin-stabilizing drug, does not show high correlations with other tubulin-interacting agents at GI50 (Supplementary Table S1); however, this may be due to the fact that the compound has much greater potency in the screen. Pemetrexed, an antifolate, has a similar pattern of growth inhibition to other antimetabolites, including floxuridine.
We found a number of agents to be inactive in the NCI60 screen at the concentrations tested. These include thalidomide, lenalidomide, aminolevulinic acid, and levamisole. It is likely that the efficacy of at least some of these drugs depends on their effect on the immune system or components of the extracellular milieu of tumors (26), which would not be detectable in a cell line screen such as the NCI60.
Discussion
As a service to the cancer research community, we undertook to screen the majority of U.S. FDA-approved anticancer drugs in the NCI60, a panel of 60 human tumor cell lines, and to make these data publicly available. The NCI60 has been used for the past two decades to screen chemicals and natural product extracts for the ability to inhibit the growth of, or to kill, cancer cells. Nearly 100,000 pure compounds and 50,000 natural product extracts have been tested, with data publicly available for pure compounds that are not covered by a confidentiality agreement, through the NCI-DTP website (4).
Although an early design hypothesis of the NCI60 screen was to identify compounds with disease specificity (i.e., compounds that might target colon cancer cells), it soon became clear that mechanistic insight into the action of novel compounds could be gained by studying the patterns of which cells responded to an agent and which were more resistant. For example, compounds that bind to tubulin have similar growth inhibition patterns regardless of which site on tubulin they bind, or whether they stabilize or destabilize microtubules. Paull et al. formalized this observation with the COMPARE algorithm (16). Dose-response curves for each cell line are converted into “end point” patterns, which represent a snapshot of the activity of the agent. Three end points are routinely calculated – GI50, i.e., the concentration at which growth is 50% of the no-drug control; TGI, total growth inhibition, the concentration where the number of cells is equal to those at time zero,when drug is added; and LC50, the concentration at which the number of viable cells is 50% of those present at time zero. These end point patterns thus make up a pattern that the COMPARE algorithm utilizes to calculate the PCC, a measure of how similar the patterns are. A correlation of 1.0 identifies a perfect match, a PCC of -1.0 denotes a perfect mirror image, whereas a PCC of 0 means there is no correlation between the two patterns. Such correlations do indeed allow one to group many of the approved drugs according to mechanism of action, as shown in Figs. 2 and 5.
All of the NCI60 datasets described herein are publicly available. The raw data for percent growth, the parametric analyses for GI50, TGI, and LC50 datasets, as well as the microarray and other molecular target characterization data are available to download, should users wish to undertake their own analyses. A variety of visualization tools are currently available on the DTP website. Both compound sensitivity and molecular target data can be searched, and resulting data displayed as a mean graph. The COMPARE algorithm can be accessed to search for NCI60 patterns similar to any starting “seed.” One can choose a compound of interest and query the database for compounds with similar patterns of activity, or for molecular targets whose pattern of expression correlates with sensitivity to an agent of interest. For instance, one can begin with a novel compound that has been tested in the screen, and run COMPARE to see if it has a similar sensitivity pattern to any agents of known mechanism, thus generating hypotheses as to the mechanism of action of the novel compound that can be tested in the laboratory.
When reviewing the results of a COMPARE analysis, there are a number of factors to consider in evaluating “hits.” How many experiments were done with the compound? Compounds with good activity and/or interesting patterns of cell line growth inhibition are generally tested at least twice. Sometimes compounds are tested in multiple concentration ranges; the values used for COMPARE are averaged for each cell line across all tests at a given concentration range, as indicated by the log of the highest concentration tested (LHICONC). Different concentration ranges may give different patterns of growth inhibition and different COMPARE results, if one of the ranges is not optimal for the calculated end point (e.g., bortezomib at the high concentrations of 10−4 mol/L and 10−6 mol/L).
In addition to the NCI60 cell line data described here, the FDA-approved agents used for these studies are available from DTP as a plated set of compounds, for use in cancer research. The agents are provided on 96-well plates as 20 μL of a 10 mmol/L solution in 100% DMSO. As new anticancer agents are approved by the U.S. FDA, DTP will add them to this set. Instructions for obtaining the Approved Oncology Drug Set, as well as ancillary information on these compounds can be found at the DTP website (17).
Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed.
Acknowledgments
We thank the members of the Developmental Therapeutics Program for providing the agents used in these studies and for making them available to cancer researchers via the Approved Oncology Drugs plated set.
Grant Support: Federal funds from the National Cancer Institute, National Institutes of Health.
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.
Footnotes
Note: Supplementary materials for this article are available at Molecular Cancer Therapeutics Online (http://mct.aacrjournals.org/).
- Received February 12, 2010.
- Accepted February 26, 2010.
- ©2010 American Association for Cancer Research.
References
Volume 9, Issue 5
