Molecular Target Class Is Predictive of In vitro Response Profile

The expanding development and use of targeted therapies for cancer treatment reflects an increasing understanding of key oncogenic pathways and how the targeted perturbation of these pathways corresponds to clinical response. Difficulties in predicting efficacy to targeted therapies is likely a consequence of the limited global knowledge of causal mechanisms for pathway deregulation (e.g., activating mutations and amplifications). Preclinical translational research studies for oncology therapies focus on determining what tumor type and genotypes are most likely to benefit from treatment. Treating selected patient populations may help maximize the potential of a therapy. The successful transition of preclinical tumor response profiles (i.e., those tumor cells having maximal efficacy) to the subsequent clinical development is now well established. Numerous specific molecular markers are now used to identify patients most likely to benefit in a clinical setting. For example, in vitro, imatinib selectively kills cells with the activated BCR-ABL gene fusion (1), whereas lapatinib preferentially inhibits proliferation of ERBB2-overexpressing cells (2). Both have achieved commercial success, benefiting patients with tumors harboring these genetic aberrations. Although inhibition of activated oncogenic pathways (normally via mutation/amplification) is thought to be the predominant mechanism of action for most novel cancer therapies, the diversity of compound selectivity for different targets obscures the logical association between compound target activity and efficacy in cancer cells. For example, the in vitro assessment of kinase inhibitory profiles of 38 unique kinase inhibitors revealed great diversity in selectivity even for those in the same molecular target “class” (3). The consequences of a single molecule inhibiting multiple kinases are complex and poorly understood in the context of clinical response. Further, the inhibition of multiple kinases not associated with the primary therapeutic target may reveal unexpected relationships in cellular response patterns between tumors. This is likely due to the existence of many more contributing cancer-promoting pathways than previously thought—the result of the unique genomic profile of each tumor. For example, comprehensive scans for somatic mutations in breast and colon cancers indicate that each tumor harbors an average of 90 nucleotide changes, ∼10% of which are predicted to be functionally relevant (4). All functionally significant somatic aberrations likely influence therapeutic response to some degree.

Although the specific “signature” of genetic lesions in each tumor likely results in a unique but predictable cellular response profile to therapeutics, to date, few have investigated patterns of cancer cell response between different therapeutics. There is some indication that predictable patterns may exist. Response of the JFCR39 tumor cell line panel to a diverse set of 130 chemicals found that a majority of compounds segregate by shared mechanisms of action (5). Early analyses of the NCI-60 tumor cell line set identified correlated patterns of proliferation response between topoisomerase inhibitors and doxorubicin (6). Identifying negative relationships is equally as important for implementing pharmacogenomics in cancer treatment. Specific breast tumors responding to 5-fluorouracil are resistant to Adriamycin and docetaxel (7). High-throughput screening of a set of tumor-derived cell lines of diverse origin is a system for studying relationships in cellular efficacy among a set of cancer therapeutics. Here, we present a meta-analysis of tumor cell line responsiveness to several general classes of targets, including (a) signal transduction pathway inhibitors, (b) mitotic inhibitors, and (c) antiangiogenics (Supplementary Table S1). When taken as a whole, response profiles are highly predictive of compound therapeutic target class. In the case of inhibitors of signaling pathways, response profiles correspond to the oncogenic cascade that it targets. Additionally, this study addresses how the relative specificity of individual therapies (e.g., a lack of specificity in kinase inhibition) affects response profiles. Finally, both individual tumor types and specific genotypes (e.g., those harboring specific mutations) seem to contribute to the degree of responsiveness to compounds of different therapeutic target class.

A total of 311 unique cancer cell lines were purchased from several vendors (American Type Culture Collection; Developmental Therapeutics Program, National Cancer Institute; German Resource Centre for Biological Material; and European Collection of Animal Cell Cultures) and grown to standard culture media recommended by the vendor (Supplementary Table S2). Typically, this included RPMI 1640 supplemented with a final concentration of 10% fetal bovine serum, 2 mmol/L glutamax, and 1 mmol/L sodium pyruvate. The cell line panel was compiled to effectively represent common tumor types (e.g., breast, lung, and colon cancers) as well as those tumor types that may be relevant in drug development. This was constrained by the availability of cell lines for certain tumor types (e.g., prostate cancers).

Nineteen compounds representing a diverse panel of cancer therapeutics were either synthesized by the Department of Medicinal Chemistry at GlaxoSmithKline or obtained commercially (Supplementary Table S1). This includes multiple compound inhibitors of the phosphatidylinositol 3-kinase (PI3K)/AKT/mammalian target of rapamycin (mTOR) pathway (n = 5), insulin-like growth factor-I receptor (IGF-IR; n = 4), and those targeting mitosis (n = 5). Additionally, several inhibitors of receptor tyrosine kinases (RTK) were included in the compound panel. Finally, data from 12 additional compounds were downloaded from resources provided by McDermott and colleagues (8). A second-tier cluster analysis was done with the 31-compound set. Cell lines common to both sets were matched by name.

Phase I. A total of 289 cell lines were seeded into 384-well microtiter plates at two densities (low and high) ranging from 300 to 3,600, depending on the rate of cell proliferation. Each density was repeated in triplicate for every cell line. After seeding, the cells were incubated at 37°C in 5% CO₂ for 24 hours. A total of 12 compounds (Supplementary Table S1, group A) suspended in DMSO were added to the cells by a Labcyte Echo550 liquid handler to give final concentrations of 10 μmol/L, 3.1640 μmol/L, 1.0010 μmol/L, 316.9 nmol/L, 100.3 nmol/L, 31.7 nmol/L, 10 nmol/L, 3.2 nmol/L, 1 nmol/L, 0.3 nmol/L, and 0. One set of DMSO controls was analyzed at the time of compound dosing to generate a time 0 data point. For the T = 0 samples, the media were aspirated, the cell nuclei were stained with 4′,6-diamidino-2-phenylindole (DAPI), and the fluorescent intensity was measured using an InCell1000 High Content Analyzer. The remaining cells were incubated at 37°C in 5% CO₂. After 72 hours, the media were aspirated and the cells were stained with DAPI and quantified by an InCell1000 High Content Analyzer.

Phase II. A CellTiter-Glo Luminescent Cell Viability Assay kit (Promega Corp.) was used to evaluate the effects on cell viability of seven additional compounds (Supplementary Table S1, group B) by measuring ATP levels corresponding to cell density. The cell lines comprising phase II, largely a subset of those included in phase I, were selected based on observations on growth rates and viability from the phase I study. In addition, those cell lines of similar genetic background (see below), which showed identical response profiles in phase I, were excluded from screening in phase II. As with phase I, cell lines were seeded in triplicate into 384-well microtiter plates at two densities, where each density represented a 2-fold difference for that cell line. A range of 10 concentrations of each compound was added to every culture as described above. After incubation for 72 hours at 37°C, 100 μL of CellTiter-Glo reagent were added and luminescence was measured with a multilabel counter (Wallac).

The luminescence of compound-treated cells was compared with the average of the 24 DMSO-treated control wells at each cell density for every concentration of compound for all triplicate wells. The percent intensity values were used in model 205 of XLfit in Microsoft Excel to calculate gIC₅₀ values for the low- and high-seeding densities (a four-parameter logistical fit; model details can be found in Supplementary Materials and Methods). Specifically, the midpoint of the growth window (the gIC₅₀) falls halfway between the number of cells at the time of compound addition (T = 0) and the growth of control cells treated with DMSO at 72 hours. The gIC₅₀ value serves as a metric for measuring the inhibition of proliferation in cancer cells. The curves for each seeding density for every cell line were manually inspected for both data quality and appropriateness of curve fitting. Where the starting data were poor, the curve and subsequent gIC₅₀ values were excluded from further analysis. Cell line seedings were considered of poor quality when the r² value of the fitted curve was <0.75. Data of poor quality were removed from all subsequent analysis.

DNA copy number alterations were calculated for each cell line using the Affymetrix 500 K “SNP chip” (Affymetrix, Inc.). DNA was extracted from each line using Mini DNeasy kit (Qiagen, Inc.) and purified. Two aliquots (250 ng each) were digested with the restriction enzyme Nsp or Sty (New England Biolabs). Digested DNA was subsequently ligated to an adaptor and amplified by PCR using Platinum Pfx DNA Polymerase (Invitrogen), yielding a product of approximately 250 to 2,000 bp. For each enzyme digest, PCR was carried out in four 100 μL aliquots, pooled, purified, quantified, normalized to 40 μg/45 μL, and fragmented with DNase I to yield a size range of approximately 25 to 200 bp. The fragmented products were labeled, denatured, and hybridized to the Affymetrix 500 K chip. On completion of hybridization, each assay was washed and stained using Affymetrix fluidics stations. Image data were acquired using the GeneChip Scanner 3000. Genome-wide estimates of DNA copy number were calculated as described by Greshock and colleagues (9). A panel of lymphoblastic cell lines (n = 20) was used to calibrate diploid for all cancer cell lines. Additionally, to ensure the unique genetic identity of each cell line, SNP chip genotypes were calculated using the BRLMM algorithm (10) and compared between all cell lines. The genetic fingerprint (i.e., each genotype) was compared for every cell line to all other cell lines in the set. Each comparison resulted in a “percent identity” that represents the genetic similarity of each cell line, where the potential values range from 0% to 100%. Cell lines having >80% identity were considered of a similar genetic origin. Additionally, the genotype profiles of this cell line set were compared with that published by the Wellcome Trust Sanger Institute's Cancer Genome Project (11). Those cell lines with similar genetic profiles are noted in Supplementary Table S2. Details about the genotype comparisons can be found in Supplementary Materials and Methods.

Transcript abundance was quantified by using the Affymetrix U133 Plus2 GeneChips in triplicate. Briefly, cell lines were plated in triplicate and lysed in Trizol (Invitrogen Life Technologies Co.). Lysates were captured with chloroform and purified using RNeasy Mini kit (Qiagen). cDNA was prepared from 5 μg of total RNA using the SuperScript Double-Stranded cDNA Synthesis kit (Invitrogen) and amplified using the Enzo BioArray High-Yield RNA Transcript Labeling kit (Enzo Biochem, Inc.). Finally, the samples were fragmented and hybridized to the HG-U133 Plus2 GeneChips, stained, and scanned according to the manufacturer's protocols. Transcript abundance was estimated by normalizing all probe signal intensities to a value of 150 using the mas5 algorithm in the Affymetrix Microarray Analysis Suite 5.0. Both gene expression and SNP chip data are publicly available (12).

For each compound, gIC₅₀s were first normalized to the overall median for that compound and transformed to a log₁₀ scale. This produced a scaled proliferation score that is a potency-independent means of comparing response profiles. All subsequent clustering analysis used Pearson's distance as a metric and was based on the average distance between nodes [using code provided by de Hoon and colleagues (13)]. Tree plotting was done with the Phylodendron software package (14).

Although all compounds elicited some degree of growth inhibition in at least a subset of the cell line panel, response profiles across the compound set were diverse (Supplementary Table S2; summary gIC₅₀ values in Table 1). For example, BEZ-235 showed relatively strong potency to induce proliferation inhibition, where 96% (188 of 196) of lines had gIC₅₀s of <500 nmol/L. In contrast, few cell lines showed such response levels to pazopanib (9 of 263; 3%). Overall, the antimitotic compound group induced comparatively high growth inhibition among compound classes. In every case, the overall median (Q₂; 50th percentile) was <150 nmol/L. In addition, the antimitotics exhibited response profiles distinct from the other compounds, where the highly responsive majority of cell lines strongly deviate from the few less responsive lines. For example, the mean gIC₅₀ for all lines <Q₃ (75th percentile) for the PLK1 inhibitor GSK461364 is 7.6 nmol/L, whereas those >Q₃ was 258 nmol/L (Fig. 1A and C). By contrast, despite a dynamic range similar to GSK461364, the dual ERBB1/2 RTK inhibitor lapatinib has few responsive lines, where the mean gIC₅₀ for those <Q₁ was 1,888 nmol/L, and just 5.2% (13 of 267) had gIC₅₀s of <150 nmol/L (Fig. 1B). This suggests that inhibiting mitosis is more uniformly effective in slowing cellular proliferation for in vitro tumor models. Conversely, relative to other classes, inhibitors of the PI3K/AKT/mTOR pathway showed more diverse potency, where Q₃ ranged from <150 nmol/L (BEZ-235) to >10,000 nmol/L (perifosine). This profound disparity is likely the result of unique targeting strategies (i.e., where in the pathway) and differential potency between inhibitors of this pathway.

Compounds of similar target classes associate by the hierarchical clustering of their median-scaled proliferation inhibition scores (Fig. 2). In other words, specific compound target classes are predictive of which specific cell lines are most responsive to increasing drug concentrations. The overall degree of correlation between compounds of similar target classes is, in part, likely a manifestation of the variable levels of growth and survival dependence each cell line has on the activation of different oncogenic pathways. This is shown by the fact that compounds of a common molecular target class cluster together due to similar subsets of cell lines showing high levels of dose-dependent response. For example, all four inhibitors of IGF-IR cluster together due to strongly associated in vitro profiles [mean Pearson's distance of 0.34 (± 0.10; n = 6) compared with an overall mean distance of 0.88 (± 0.19; n = 145); Fig. 2, cluster A]. In contrast, although four of five drugs targeting the PI3K/AKT/mTOR pathway do cluster together (Fig. 2, cluster B), their association is considerably weaker than that of IGF-IR inhibitors (mean distance = 0.63 ± 0.15; n = 10). This likely reflects the more diverse means of targeting this pathway by this set of compounds or that each of these targets (PI3K/AKT/mTOR) has slightly unique signaling outcomes. Selective inhibitors of AKT and mTOR target this pathway downstream of those targeting PI3K directly. The two pan-isoform PI3K inhibitors in this study, BEZ-235 and GSK1059615, have notable and direct inhibition of mTOR kinase, which makes the relationships of their response profiles with other pathway inhibitors more complex. The correlation between these inhibitors was not driven by structural similarity. For example, two more strongly correlated members of this panel of inhibitors, the AKT inhibitor GSK690693 and the PI3K inhibitor GSK1059615 (Pearson's distance = 0.54), are structurally distinct (Supplementary Table S3). The mitosis-inhibiting compounds (PLK/AURK/CENPE/tubulin) also strongly correlate, where five of five members of cluster C (Fig. 2) are of this class. The mean Pearson's distance for the pairwise comparison of these is 0.56 (± 0.11; n = 10). The specific mitotic targets of these compounds are largely unique, and their clustering is consistent with previous observations that the biological process targeted (mitosis in this case) can drive similarities in cellular response (15). The few cell lines whose proliferation is not affected by treatment with an antimitotic help define this compound cluster. For example, the high correlation of those lines >Q₃ contributes to the correlation between GSK461364 and paclitaxel (Fig. 1C), a statistically significant relationship by χ² analysis when binning cell lines by quartile (γ = 0.34, P = 0.002; Fig. 1D; ref. 16). It is likely that the medium conditions conducive to high rates of cellular proliferation produce an artificially high response rate to those compounds targeting mitosis (17).

Clustering analysis also suggests logical relationships between the response profiles of the compounds not in one of these major classes. For example, the closest member to the PI3K/AKT/mTOR cluster is lapatinib, an inhibitor of the ERBB1/2 RTKs that signal through PI3K. Genomic aberrations to ERBB1/2, which have been correlated with cellular response to lapatinib (2), can activate PI3K signaling. Conversely, direct activation of PI3K via mutations to PIK3CA confers resistance to trastuzumab, another upstream inhibitor of ERBB2 (18). In other words, some, but not all, cell lines that respond to PI3K/AKT inhibition do not respond to upstream inhibition of the ERBB1/2 RTKs. The clustering of the mitogen-activated protein kinase (MAPK)/extracellular signal-regulated kinase (ERK) kinase (MEK) inhibitor GSK1120212 remains separate from those compounds targeting the PI3K/AKT/mTOR pathway and lapatinib. This relationship is likely the result of the high importance of activated RAS/RAF in those cells that are more responsive to MEK inhibition (19).

The reclustering of the 19 compounds from the initial study with additional proliferation inhibition data from 12 compounds presented by McDermott and colleagues (8) further illustrates that therapeutic target selectivity is predictive of cell line response profiles (Fig. 3). For example, the node containing the ERBB1 inhibitor erlotinib and the ERBB1/2 inhibitor HKI-272 cluster with the other RTK inhibitor lapatinib (cluster D) in this histologically diverse panel of cell lines (n = 117). In addition, the lone additional mitotic inhibitor MK0457, which targets Aurora kinases (AURK), clusters with the five other mitotic inhibitors. Similarly, the c-MET inhibitor PHA665752 clusters with GSK1363089 (foretinib), whereas the RAF inhibitor AZ628 clusters with the MEK inhibitor GSK1120212. Although these results show that the primary target is the dominant driver of gross cell line response profiles, the potent inhibition of additional targets and degree of selectivity play a role in these relationships. Only two of four compounds known to potently inhibit vascular endothelial growth factor (VEGF) components (pazopanib, foretinib, sunitinib, and sorafenib) cluster together (Fig. 3, cluster E). Sunitinib, a potent inhibitor of VEGF receptor (VEGFR) and platelet-derived growth factor receptors (20), clusters with pazopanib. Sorafenib, a potent inhibitor of VEGFR, RAF, and numerous other kinases (21), does not cluster with other inhibitors of VEGFR. This is not surprising, as a direct comparison of these three inhibitors of tumor vasculature reveals that each potently inhibits a unique set of kinases known to be involved in tumorigenesis and progression (22). It is likely that inhibiting these additional kinases drives proliferation response, as cell lines do not serve as faithful models of tumor vasculature in two-dimensional cell culture. In the case of sorafenib, the inhibition of kinases unique to this compound (compared with that of foretinib and sunitinib) is more critical in determining cellular response in vitro. In addition, although sunitinib and sorafenib show comparable benefit as second-line treatment of cytokine-refractory kidney cancers (23), it is not known if the same patients would benefit from each therapy.

Relationships between cluster groups also show rational patterns of cell line response that are not directly related to the target but rather the cascading of cellular signaling. For example, inhibitors of the RTKs ERBB1/2, in addition to the SRC inhibitor AZD0530 [also shown to selectively inhibit growth of epidermal growth factor receptor (EGFR)–activated cells (8)], cluster with inhibitors of PI3K/AKT/mTOR. Although it is established that the RTKs inhibited by these compounds (ERBB1/2) modulate PI3K/AKT/mTOR in numerous human cancers, the relationship between PI3K/AKT/mTOR activity and clinical response to RTK inhibition is complex and only partially understood. For instance, activating mutations of the PIK3CA gene confer resistance to EGFR inhibition (24) while correlating with sensitivity to compounds targeting PI3K directly (25).

The observed compound class clustering associations are consistent with previous pharmacogenomic studies, showing that cell line response to therapeutics is dependent on both genotype and cellular origin (8). When considering the response data for cells of breast, colon, lung, and hematopoietic origin in the original panel of 19 compounds (those tissue types represented by at least 15 cell lines), inhibitors of the PI3K/AKT/mTOR pathway show increased selectivity for inhibiting the growth of breast cells (Fig. 4). Perifosine, which exhibited little growth inhibition in this cell line panel, was the lone exception. The selective growth inhibition in breast cancer cells by this class of compounds is expected, as this pathway is shown to be frequently activated by mutations of PIK3CA (26) and is of importance in driving breast cancer cell growth (27). Hence, breast cancers may represent a favorable clinical target for this emerging class of therapeutics. Similarly, four of four IGF-IR inhibitors showed uniform selectivity of colon cancer cell lines when compared with the other tumor types. The preferential efficacy of IGF-IR inhibition in colon cancer cells is consistent with the frequently observed expression of the IGF-IR in colon cancers when compared with other tumor types, where it is not promoting tumorigenicity (28).

Overexpression of estrogen receptor (ER), which stimulates cell division and DNA replication in ∼70% of breast cancers, serves as a pathologic marker for diagnosis and treatment. The discrimination of the luminal A subtype from other molecular subtypes of breast cancers is defined largely by overexpression of ER and amplification of ERBB2 (29). These markers have profound implications for breast cancer prognosis and survival (18) and correlate with cellular response to therapeutics such as dasatinib (30) and lapatinib (2). It is then not surprising that ER-overexpressing breast cancer cell lines segregate from those with low ER expression (P = 0.010, Fisher's exact test; Fig. 5). This relationship is even stronger when considering molecular classifications defined by Sorlie and colleagues (29), where six of seven (86%) and nine of nine (100%) cell lines of basal and luminal subtypes cluster with their respective primary tumor subtypes (P = 0.0009). Notably, luminal cells preferentially respond to the AKT inhibitor GSK690693 and the PI3K inhibitor GSK1069615, whereas the MEK inhibitor GSK1120212 selectively inhibits proliferation in basal breast cancers. This result is not surprising because PI3K activation can mediate resistance to MEK inhibition (31). Although, collectively, this suggests a strongly divergent clinical path for targeting these pathways in breast cancers, the coinhibition of MAPK and PI3K may be particularly efficacious in the basal subtype (32). Mutations in PIK3CA and PTEN do not seem to associate with compound clusters (P > 0.05), suggesting that these alterations are of secondary importance to ER and ERBB2 in determining what cell lines are more responsive across this diverse panel of compounds. Similarly, when analyzing tumor types with frequent KRAS mutations (colon and pancreatic cancers), those cell lines harboring mutant KRAS tended to cluster together (P = 0.14, Fisher's exact test; Supplementary Fig. S1). This trend is even more profound when distinguishing those harboring KRAS mutations with wild-type PIK3CA (P = 0.008, Fisher's exact test). The unique sensitivity profile of KRAS-mutated cell lines is an expected result, as mutations to this gene correlate with clinical response patterns, such as resistance to inhibitors of EGFR (33). The fact that not all KRAS-mutant tumors cluster together is also concordant with previous observations that mutations that co-occur with KRAS can drive cellular response to therapeutics, such as the association of cooperating LKB1/KRAS mutations with more exquisite response to MEK inhibition (34).

This study shows that although kinase selectivity of cancer therapeutics can vary greatly (3), the predominant targets of these therapies play a major role in determining inhibition of cell proliferation in a dose-dependent manner in vitro. The similarity in concentration-response profiles among multiple inhibitors targeting a common oncogenic pathway emphasizes how inhibition of that specific pathway is a key cellular response mechanism. In addition, the unique but consistent profiles from distinct therapeutic target classes (e.g., PI3K/AKT/mTOR versus IGF-IR) are suggestive of divergent clinical strategies for these compounds. Large panels of tumor-derived cell lines show response affinities for molecular target classes with specific tumor types (as was the cases with PI3K/AKT/mTOR and IGF-IR inhibitors for breast and colon cancer cells, respectively). Further, cellular response profiles among classes show genotype specificity, exemplified by the congregation of luminal/basal breast cancer cells across compounds of diverse target classes. These results also show that cell culture is a largely faithful model of primary tumors with respect to drug response, particularly for those compounds targeting signal transduction pathways. In contrast, hyperphysiologic growth conditions may hinder the accurate modeling of drug response of some compound target classes, such as those designed to inhibit mitosis. This notion may become increasingly important in the future of cancer drug development as target classes expand to include those selectively obstructing tumor metabolism because specific cell culture conditions can drastically affect metabolic activity (35). Collectively, the high-throughput screening of cancer therapeutics across a histologically and genetically varied cell line population can simulate clinical response profiles and identify both tumor types and genotypes that may benefit most from specific treatment regimens.

All authors are GlaxoSmithKline stock owners.

We thank Adam Butler, Jon Teague, and Mike Stratton (Wellcome Trust Sanger Institute) for help with cell line genotype matching and mutation data management.

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.