Three-kinase inhibitor combination recreates multipathway effects of a geldanamycin analogue on hepatocellular carcinoma cell death

Introduction

The fairly limited success of many targeted cancer therapeutics when used as single-agent treatments presents a challenging problem that has motivated studies seeking to stratify failure-versus-success categories (1) or to combine targeted therapeutics with traditional chemotherapeutic regimens (2). An alternative avenue generating growing interest is a new class of compounds known as multitarget drugs (3). These compounds aim to improve therapeutic efficacy by producing a combined inhibition of diverse regulatory pathways that are important for cancer cell proliferation and survival.

At one end of the multitarget drug spectrum lies a class of Hsp90 inhibitors, such as 17-allylamino-17-demethoxygeldanamycin (17AAG), which were derived from the tumoricidal natural product geldanamycin (4). Hsp90 is a vital chaperone that, relative to other chaperones, interacts with a select but critical subset of cellular proteins, including nuclear hormone receptors and components of signal transduction cascades (5). Hsp90 inhibition seems to offer a promising anticancer strategy (6) but, compared with more traditional targeted therapeutics, an exceptionally pleiotropic effect. In a recent genome-wide study in Saccharomyces cerevisiae, Hsp90 was found to interact with ∼10% of the open reading frames investigated (7).

The nature of the effects of 17AAG in tumor cells relative to nontransformed cells seems to derive from the enhanced binding of 17AAG to Hsp90 in the former. Tumor cell Hsp90 is found more highly resident in multichaperone complexes with high rates of ATPase activity, constituting a distinctive molecular state exhibiting an ∼10-fold greater binding affinity for 17AAG compared with that in the normal cells (8). Hsp90 inhibition in a variety of tumor cell lines has been shown to affect the levels and/or activity of ErbB family receptors, Src, Ras, Raf, AKT, IκB kinase (IKK), Janus kinase, Her2, p53, RIP, and cell cycle regulators, as well as increasing the levels of the antiapoptotic chaperone Hsp70 (5, 9–14). In addition to these long-term effects on client protein levels and their associated downstream signaling effects, short-term effects of Hsp90 inhibitors have also been observed. Geldanamycin treatment has been found to yield increased phosphorylation of AKT in myocytes (15), increased PKR activity and phosphorylation of eIF-2α in HeLa cells (16), and transient early signaling in the Src, AKT, and extracellular signal-regulated kinase (ERK) pathways in both MCF-7 and COS7 cells (17). However, the contributions of these various pathway effects to cell death responses are unclear, and the manner by which these effects integrate to affect cell behavior is even more difficult to ascertain.

Our goal is to offer an effective approach based on systematically extracting the critical effects of multitarget compounds, which could enable the rational recreation of their effectiveness via designed combinations of more selectively targeted drugs. Because Hsp90 inhibition elicits contradictory effects, the degradation of proteins involved in cellular survival as well as an increase in Hsp70 levels, we reason that a rational recreation of its proapoptotic effects may potentially increase efficacy. Our approach is motivated by recent successes in computational characterization of the effects of diverse drug-induced perturbations governing cellular phenotypes (18, 19). Data-driven computational modeling techniques, such as principal components analysis (PCA) and partial least-square regression, seek to find key vectors representing signal combinations that contain the most vital information for predicting—at least in correlative fashion—cell responses to various stimuli (19).

We test this hypothesis on quantitative experimental measurement of multiple phosphoprotein signaling pathways altered by 17AAG in a set of hepatocellular carcinoma (HCC) cell lines. A PCA of time course data for seven kinase signals reveals that early effects of 17AAG on the AKT, IKK, and signal transducer and activator of transcription 3 (STAT3) pathways are predominantly critical for clustering cell-specific apoptotic death responses among the Hep3B, Huh7, and FOCUS lines. Whereas individual inhibition of each of these three pathways had little effect on the cell responses, combining all three kinase inhibitors rendered the 17AAG-resistant Huh7 line responsive and had greater efficacy than 17AAG itself in two of our three cell lines.

Results

Conventional Markers of 17AAG Action Cannot Account for the Differential Cell Death Responses in Huh7 and Hep3B Cells

The effect of 17AAG treatment on cell death in HCC cell lines was assessed by treating Huh7 and Hep3B with 17AAG at 0.1 and 1 μmol/L for 48 hours. A representative gating for a population staining double positive for active (cleaved) caspase-3 and cleaved PARP was assessed via flow cytometry (Fig. 1A). Double-positive cells represent the population of cells within a well that are undergoing apoptosis. The same gating thresholds were used for all of the samples from a given cell line and are based on the negative control. No single positive populations were noted at 48 hours during the duration of our study. Three samples per cell line/condition were averaged using the values obtained from the plots in Fig. 1A and are plotted in Fig. 1B. Hep3B cells were found to be sensitive to 1 μmol/L 17AAG–induced cell death, whereas Huh7 cells were found to be resistant (Fig. 1B). We then hypothesized that a difference in the levels of phospho-AKT (Ser⁴⁷³), total AKT, phospho-ERK1/2 (Thr²⁰²/Tyr²⁰⁴), and phospho–IκB-α (Ser³²/Ser³⁶) 24 hours after the addition of 17AAG might account for the difference in phenotype (Fig. 1C). Lysates were collected both before and 24 hours after treatment with 17AAG. The relative decrease in these four measurements at 24 hours is identical in sensitive and resistant cell lines. These late-time measurements indicate that a simple explanation about the long-term signaling degradation arising from 17AAG treatment is not able to explain the disparate death responses of Hep3B and Huh7 cells.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

A plot for the cleaved caspase-3/cleaved PARP double-positive population at 48 h measures the susceptibility of Hep3B and the resistance of Huh7 cells to 17AAG. Typical measurements of 24-h signaling degradation in response to 17AAG fail to correlate with this distinction. A, a representative flow cytometry scatter plot depicts Hep3B-fixed and Hep3B-permeabilized cells that are stained with antibodies for active cleaved caspase-3 and cleaved PARP at 48 h after treatment with 1 μmol/L 17AAG or 0.1% methanol vehicle control. The double-positive population is denoted by the gating in the upper right-hand corner. Different gatings were used for Huh7 and Hep3B but the scatter plots looked very similar. B, the average size of the population of three replicates (±SE) of Hep3B and Huh7 cells as gated in A represents the percentage of double-positive apoptotic cells at 48 h ± SE. C, late time signals do not correlate with 17AAG susceptibility. Measured at 24 h, the fold change of the mean of duplicates ± SE, treated with 1 μmol/L 17AAG, is normalized to a 0.1% methanol control. Phospho-AKT (Ser⁴⁷³), total AKT, phospho-ERK1/2 (Thr²⁰²/Tyr²⁰⁴), and phospho–IκB-α (Ser³²/Ser³⁶) were measured by a bead-based Bio-Rad phosphoprotein (Bioplex) assay.

A Dynamic Multipathway Analysis Can Distinguish Huh7 Resistance and Hep3B Sensitivity to 17AAG

Because the degradation of phosphatidylinositol 3-kinase (PI3K)-AKT, Ras/Raf/MEK/ERK, and IKK-NF-κB pathway signaling at 24 hours failed to distinguish Huh7 from Hep3B, we proposed that a dynamic study including measurements of the shorter-term effects of 17AAG treatment might yield information that could distinguish the differential response behaviors of these cell lines. We thus measured the levels of the same signals at 0, 1, 2, 4, and 24 hours after treatment in both the Huh7 and Hep3B cells. These time points were initially chosen based on the literature reports of early and late changes in phosphorylation levels in response to 17AAG (17). These measurements were normalized to an untreated control for each cell line and are plotted in Fig. 2A as fold change induced by 17AAG. Both Huh7 and Hep3B exhibited transient early phospho-JNK activation. However, relatively stronger transient activation and/or relatively weaker degradation were seen for most other signals (phospho-AKT, phospho–IκB-α, phospho–GSK-3α/β, phospho-ERK1/2, phospho-p38, and phospho-STAT3) in Huh7 cells compared with Hep3B cells (Fig. 2A). Strikingly, the integrated signal changes over the first 4 hours after 17AAG treatment are higher across many components in Huh7 cells than in Hep3B cells (Fig. 2B). To more effectively interpret these data, we used PCA, a technique that models multivariate data in terms of key combinations of measurements exhibiting major covariation in the data set. These combinations can be viewed geometrically as “directions” (or vectors) in the “space” of the signaling measurements defined by axes representing a combination of the measured time points and the integral metric. The direction of greatest covariation (an axis representing the time points that are most similar), which incorporates the most important measurement combinations, is the principal component 1 (PC1); principal component 2 (PC2) then incorporates the next most important measurement combinations best capturing the residual covariation, and so on. The principal components collapse the original multidimensional data into a reduced number of axes representing the most important underlying variables. In our case viewing a “loadings plot” (data summarized in Fig. 3C), the six dimensional data sets (1 hour, 2 hours, 4 hours, 24 hours, and early/late integral metrics) can be comprehended in two-dimensional principal components space as early signaling (representing primarily 1 hour, 2 hours, and the early integral data; PC1) and late signaling(where the 24-hour data project most strongly; PC2). These underlying variables are just combinations of the original variables that best comprehend the data within the entire signaling measurement “vector space” (Supplementary Fig. S1 offers an illustrative tutorial). Our model with two principal components effectively captures 91% of the cumulative variance. Our cutoff was readily established by the major drop in the variance captured by the third component (Fig. 3A). Furthermore, a resampling of our data set determined that our analysis captures significantly (P = 0.0002) greater variance than one can capture with 1,000 two-component models generated by randomly sampling the column index of the rows in our data set (Fig. 3B). Our analysis visually distinguishes the behavior of both Huh7 and Hep3B cells. A so-called “scores plot,” in which the various signaling measurements are projected on the principal components axes, visualizes distinct regions where Huh7 signals reside differentially with respect to Hep3B signals (Fig. 3C). These differences are quantified in Fig. 3D, in terms of “contributions vectors,” which quantify the disparities between cell line–specific signals (in units of SD) relative to the average distance between all of the signals of the distinct cell lines. The calculated contributions vectors reveal that of the 10 signaling measurements made at each of 5 time points, the most important differences between the drug-resistant Huh7 and drug-sensitive Hep3B lie in the early time points of four signals: phospho-AKT, phospho–IκB-α, phospho-STAT3, and phospho-p38. Because the experimental conditions for the elucidation of the signaling effects grow very rapidly with the size of the hit set, the phospho-p38 signaling measurements were excluded from further analysis because they reside as an outlier beyond the region of Hotelling's criterion (23) and presented no clear biological hypothesis.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

A dynamic signaling time course reveals differences between Hep3B and Huh7. A, a signaling time course in Hep3B and Huh7 cells depicts mean fold changes in phosphoprotein signaling in response to 1 μmol/L 17AAG relative to the vehicle only (0.1% methanol) control. B, the integral (discussed in Materials and Methods) from 0 to 4 h of 17AAG-induced signaling shows large cumulative differences in early signaling between Huh7 and Hep3B cells.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

A PCA reduces data complexity and provides testable hypotheses. A, percentage of the total variance captured by a model consisting of one, two, or three principal components. There is a marginal increase in the benefit of including principal component 3, indicating an obvious cutoff criterion. B, percentage of total variance explained by two-component models that are built on 1,000 perturbed data matrices (see Materials and Methods). A histogram plots the data from the empirical 1,000 matrix sampling. A normal distribution was fit to the histogram data, and our model fit was calculated to have a highly significant P value of 0.0002. C, a PCA yielded a two-component model that accurately explains 91% of the cumulative variance. PC1 strongly captures variation at early time points and in the 0- to 4-h integral metric. PC2 captures variation at 24 h. These results from the loadings plot are summarized in the upper right-hand corner of the plot. The ellipse represents Hotelling's outlier criteria at a 95% confidence level. The original signaling measurements are plotted in the principal components space. PC1 visually seems to capture cell line variation. The colored ellipses are simply a visualization tool used to bring the readers attention to the distinct clusters in the scores plot. D, contributions vectors describe how signals vary in principal components space and are derived computationally by measuring the latent variable distance in terms of the Euclidean distance of the measured variables between Huh7 and Hep3B for a given signal and then comparing that distance to an average distance of all the distances between the cell lines (in units of SD). This plot asks the question of how distinct are two signals between Huh7 and Hep3B cells relative to the average distance between lines. Quantitatively, D affirms qualitative observations made in C.

The Principal Components Model Accurately Maps Another 17AAG-Sensitive HCC Cell Line, FOCUS, to the Same Region as Hep3B

To test our PCA, we used our previous approach in a third HCC cell line. A signaling time course for FOCUS cells, analogous to Fig. 2A, shows signal degradation and no transient activation (Fig. 4B). Although certain signaling network differences between FOCUS and Hep3B signaling are evident by inspection, our PCA shows that the FOCUS cells cluster in the same region as the similarly 17AAG-sensitive Hep3B cells (Fig. 4C). Indeed, we found that FOCUS cells are sensitive to 1 μmol/L 17AAG–induced cell death at 48 hours (Fig. 4A). This result shows that the principal components capture critical signal combinations associated with these cell death-versus-survival outcomes.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Adding signaling data from another HCC cell line (FOCUS) to the PCA correctly clusters FOCUS cells with sensitive Hep3B cells. A, mean of triplicate measurements of the percentage of double-positive FOCUS cells in response to 1 μmol/L 17AAG at 48 h. The gating strategy was the same as in Fig. 1A. B, signaling time course of FOCUS cells in response to 1 μmol/L 17AAG. Time point measurements are represented as mean signaling fold change relative to vehicle only controls. C, a principal components scores plot, as in Fig. 3, correctly classifies FOCUS cell phosphoprotein signaling measurements as homologous to 17AAG-sensitive Hep3B cells in the two-component model.

A Drug Combination Pretreatment Based on Inhibition of Three Key Nodes Partially Sensitizes Huh7 to 17AAG

In Fig. 3D, we identified four signals (phospho-AKT, phospho–IκB-α, phospho-STAT3, and phospho-p38) to be most important in distinguishing Huh7 from Hep3B. We proposed to test whether a combination of drugs inhibiting these key signals in particular could recreate the 17AAG treatment responses. We selected phospho-AKT, phospho-STAT3, and phospho–IκB-α for this purpose, omitting phospho-p38 due to its notification by Hotelling's outlier criteria at a 95% confidence level (23). Pretreatment of Huh7 cells for 12 hours with targeted inhibitors of phospho-AKT (PI3K inhibitor, PI103, 5 μmol/L), phospho-STAT3 (JAK inhibitor 1, pyridone 6, 3 μmol/L; Calbiochem), and phospho–IκB-α (IKK inhibitor, BMS-345541, 15 μmol/L) signaling failed to sensitize Huh7 cells to 1 μmol/L 17AAG at 48 hours (Fig. 5A). However, a combination pretreatment using all three inhibitors together sensitized (P < 0.0001) Huh7 cells to 17AAG-induced cell death (Fig. 5A). Moreover, this inhibitor combination induced Huh7 cell death to a significantly greater (P < 0.0005) degree than did 17AAG (Fig. 5A). To control for the possibility that any combination of three inhibitors can potentiate 17AAG-induced cell death, inhibition of one critical signal (phospho-STAT3) was combined with inhibition of two noncritical signals (phospho-JNK and phospho-ERK; Fig. 3D). This control drug combination, comprising 3 μmol/L of the JAK inhibitor, 10 μmol/L of a MEK inhibitor (PD98059), and 10 μmol/L of a JNK inhibitor (SP600125), failed to sensitize Huh7 to 17AAG or induce cell death on its own at 48 hours (Supplementary Fig. S2).

• Download figure
• Open in new tab
• Download powerpoint

Figure 5.

Where inhibition of singular nodes fails, pretreatment with a combination therapy sensitizes Huh7 cells to 17AAG and works as well or better than 17AAG in all cell lines tested. A, double-positive populations of Huh7 cells (mean of triplicate ± SE as gated in Fig. 1A) in response to different treatments at 48 h. The IKK inhibitor BMS-345541 was used at 15 μmol/L, the PI3K inhibitor PI103 was used at 5 μmol/L, and the JAK inhibitor JAK inhibitor 1, pyridone 6, was used at 3 μmol/L. The P values were obtained by a t test. B, double-positive populations for the listed treatments of Hep3B and FOCUS cells at 48 h. Columns, mean of triplicates; bars, SE. The P values were obtained by a t test. The drug combination is the same as in A.

Decoupling Proapoptotic Effects from the Overall Context of 17AAG Treatment Creates a Three-Kinase Inhibitor Combination That Works as Well or Better than 17AAG in All Cell Lines

17AAG is known to have both proapoptotic and antiapoptotic effects. Assuming that 17AAG induces cell death through a combination of the proapoptotic consequences of the degradation of key signaling pathway nodes, we hypothesized that the previously identified critical nodes distinguishing sensitive and resistant cells could be key sites for selectively targeted drug contributions in sensitive cells. To test this idea, FOCUS and Hep3B cells were pretreated with the three-kinase drug combination, or control vehicle, for 12 hours before addition of 1 μmol/L 17AAG. The drug combination enhanced 17AAG-induced cell death significantly (P = 0.050 and P < 0.0001, respectively) in both cases (Fig. 5B). Yet, more strikingly, the drug combination by itself induced cell death as strongly—or more so—than 17AAG in all HCC cell lines. FOCUS cells showed no significant difference between 17AAG and the drug combination–induced cell death, whereas in Hep3B and Huh7 cells the drug combination at 48 hours worked significantly better than 17AAG (Fig. 5A and B).

Decoupling the Proapoptotic Effects from Their 17AAG Context Allows for Cell Line–Specific Tuning of the Therapeutic Strategy

After finding a drug combination that works as well or better than 17AAG, we wanted to test two hypotheses. First, that the decoupling of a specific pathway effect from the context of 17AAG allows for cell line–specific therapeutic strategies, and second, that the presence of additivity or synergy correlates with cell line–specific differences in the combination inhibitor efficacy relative to 17AAG. To address these questions, we undertook a set of experiments in a JAK inhibitor background, across a 4 × 4 matrix of PI3K and IKK inhibitor concentration combinations, covering four concentrations of each (0, low, medium, and high; Fig. 6). Inspection of the cell death responses reported in this matrix indicated distinct responses for the three cell lines tested. FOCUS cells seemed to primarily be sensitive to the IKK inhibitor in the JAK inhibitor background, whereas Hep3B and Huh7 cells exhibited sensitivity to both IKK and PI3K inhibition (Fig. 6A). Calculation of the fold increase in cell death of a double-drugged entry in the 4 × 4 dosing matrix relative to linear additivity shows that only Huh7 cells exhibit a drug combination synergy (Fig. 6B). This finding of multidrug nonlinear synergy in the Huh7 and multidrug additivity in Hep3B but not FOCUS cells correlates with our observation that the kinase inhibitor combination could essentially duplicate the 17AAG effects on cell death in the latter, whereas it could exceed the 17AAG effects on cell death in the former two.

• Download figure
• Open in new tab
• Download powerpoint

Figure 6.

FOCUS, Huh7, and Hep3B cells exhibit different amounts of multinode effects in a JAK inhibitor background. A, a matrix of the mean of duplicate measurements of the percentage of double-positive cells in response to varying concentrations of IKK inhibitor (IKKi) and PI3K inhibitor (PI3Ki) in a JAK (3 μmol/L) inhibitor background. The compounds are the same as in Fig. 4. Concentrations are as follows: IKK inhibitor: high (H), 20 μmol/L; medium (M), 6.66 μmol/L; low (L), 2.22 μmol/L; PI3K inhibitor: high, 5 μmol/L; medium, 1.66 μmol/L; low, 0.55 μmol/L. B, the synergy plots display the fold deviation from predictions based on the assumption of additivity. Briefly, additivity predictions for any double-drugged entry (i,j) were calculated by adding the ith row of the first column (i.e., the singular IKK effect at that concentration) to the first row, jth column (i.e., the singular PI3K effect at that concentration). Then, the measured value at the double-drugged entry (i,j) was divided by the additivity prediction. This created a metric that describes synergy versus additivity.

Discussion

Inhibition of the Hsp90 chaperone represents a prominent example of a highly complex multitarget therapeutic approach to cancer (5), and studies have explored the myriad of regulatory pathway alterations associated with its effects (5, 9–17). Although these studies have unveiled short-term and/or long-term changes in levels and/or activities of numerous cellular components with opposing effects on the downstream phenotype, there is currently inadequate understanding of pathway alterations that can recapitulate drug efficacy. Our effort here aims to gain quantitative understanding of the key kinase signals underlying the effects of the geldanamycin analogue 17AAG on differential apoptotic death responses of three HCC cell lines: Hep3B, FOCUS, and Huh7. Our studies centered on seven canonical kinase signaling effects, phospho-AKT (Ser⁴⁷³), phospho-ERK1/2 (Thr²⁰²/Tyr²⁰⁴), phospho-STAT3 (Tyr⁷⁰⁵), phospho-p38 (Thr¹⁸⁰/Tyr¹⁸²), phospho-JNK (Thr¹⁸³/Tyr¹⁸⁵), phospho–GSK-3α/β (Ser²¹/Ser⁹), and phospho–IκB-α (Ser³²), with a goal of ascertaining whether a quantitative combination of a particular subset of these signals might be especially critical in clustering the HCC line responses to 17AAG.

We found that a PCA-based model consisting of two principal components explains 91% of the cumulative variance captured by the model. This variation can be classified as early variation (PC1) and late variation (PC2; Fig. 3). All cell lines and signals inhabit the same principal components space, implying that there is a general homology of response. This overall homology stresses that the drug is operating similarly in all cell lines tested and that the phenotypic responses are not likely to be based on the metabolic inactivation of 17AAG or a rapid cellular efflux. The early variation axis (PC1) best separates the cell lines with disparate responses, suggesting that this previously documented early signaling phenomena (17) can be correlated with a cellular phenotype.

The PCA-based visualization accurately maps the sensitive FOCUS cell line to the same region as Hep3B cells (Fig. 4). This greatly improves our confidence that the variation captured in two-dimensional principal components space is relevant to the disparate cell phenotypes and that our analysis has captured important biological information. The success of our PCA mappings suggests that data-driven modeling can effectively reduce the dimensionality of multitarget perturbations and provide a compact, easy-to-interpret analytic method for determining the most phenotypically relevant parameters for further investigations.

Previous data-driven modeling efforts (24, 25) have shown that a systematic signaling analysis of growth factor–induced and cytokine-induced cellular phenotypes captures enough signaling variation that the phenotypic effects of drug perturbations can be predicted a priori. Kumar et al. (18) extended this analysis to show that multiple signals measuring the off-target effects of inhibitors are necessary to predict drug effects on epidermal growth factor–induced migration. Our quantitative analysis of 17AAG leads to testable hypotheses about the targeting of early IKK, PI3K, and JAK-STAT inhibition in 17AAG-induced cell death (Fig. 2). The individual targeted pharmacologic pretreatment of these three survival pathways fails to abrogate the resistance phenotype, but the combined pretreatment is able to synergistically sensitize Huh7 to 17AAG (Fig. 5A). This evidence suggests that cumulative early action of the IKK, PI3K, and JAK-STAT pathway families may be responsible for a portion of the drug resistance phenotype of Huh7.

Although the combination therapy potentiates 17AAG-induced cell death in all cell lines tested, the most striking consequence of our study is the ability of combination therapy alone to recapitulate the cell death effect of 17AAG in a manner that is equal to or greater than the singular effect of 17AAG. This effect has two distinct modalities. In Huh7 and Hep3B cells, the combination treatment works approximately 5- and 3-fold better, respectively, than 17AAG alone. In contrast, the combination treatment is of similar efficacy to 17AAG alone in FOCUS cells (Fig. 5). These distinct modalities can be correlated with the spectrum of multiple drug interactions in the combination therapy (Fig. 6). We speculate that these correlations may suggest a basis for future work on the cell-specific nature of both the proapoptotic/antiapoptotic function of 17AAG and how synergy between nodes may play a role in the efficacy of multidrug combinations and multitarget inhibitors.

By using principal components modeling of a drug perturbation to guide the recreation of a more specific therapeutic strategy, we have shown that we can not only recreate the effect of a multitarget drug but that the recreated effect can work better than the original drug. The existence of differences in the multidrug interactions (Fig. 6) underscores the value of recreating multitarget effects. In decoupling the effects from the context of 17AAG, we find that the potential exists to allow for the individual adjustment of the elements of a combination therapy.

Our suggestion that a specified combination therapy may produce a less generally pleiotropic effect than the multitarget compound from which it has been rebuilt arises from comparative consideration of the diversity of the cellular effects of the initial multitarget compound along with the off-target effects of the combination therapy. In our particular case, which is certainly a fairly extreme example, Hsp90 in S. cerevisiae is known to interact with ∼10% of the open reading frames examined in a recent study (7). Although a similar analysis is not available for the three compounds of our combination therapy, it seems plausible that our combination therapy is less broadly cross-reactive for several reasons. A systematic study of 317 kinases found that for PI103, ∼10 cellular kinases had IC₅₀s of <10 μmol/L in vitro (26). BMS-345541 is an allosteric inhibitor of IKK and failed to significantly inhibit a panel of 15 related kinases at concentrations as high as 100 μmol/L (27). Our Janus kinase inhibitor, pyridone 6, although less well tested than the other two, fails to inhibit the proliferation of cell lines that do not harbor activated JAK-STAT signaling (28). Thus, in spite of the inevitable existence of some off-target effects, it seems highly unlikely that the magnitude of these effects could approach the pleiotropy of 17AAG. Therefore, the generalizability of this approach to other forms of multitarget compounds will likely prove increasingly practical as more specific inhibitors become available.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.