Introduction

Many small-molecule compounds are used as inhibitors of cellular signaling pathways and therapeutic agents1,2,3. In both basic research and clinical settings, elucidating the target selectivity of such compounds is critical for predicting and interpreting their effects4,5,6,7. A library of kinases, for example, might be useful for measuring the effects of compounds on kinase activities and identifying the target kinase of each compound4,8,9,10. Through such approaches, it has become clear that most compounds, including many drugs in clinical use, have multiple targets. Protein libraries make it possible to screen many proteins simultaneously, but the number of proteins available in such systems is still limited relative to the diversity of proteins within living cells. Consequently, it is possible that a given compound of interest may have an unexpected target inside cells. If an as yet unknown protein is revealed as a new target, such information could explain a compound’s side effects or encourage repositioning of the compound as a treatment for other diseases11,12.

In this study, I focused on epidermal growth factor receptor (EGFR), a prototypical receptor tyrosine kinase (RTK), because this protein has been extensively investigated as an important target of small-molecule compounds in both basic and clinical research13,14,15. Inhibitors of EGFR tyrosine kinase used in clinical practice include gefitinib, erlotinib and afatinib, which are used in therapy against non-small cell lung cancers (NSCLCs) harboring EGFR mutations16,17,18. In addition to direct inhibitors of EGFR itself, compounds that affect EGFR signaling components such as K-Ras, MEK1 and PI3KCA are also candidate therapeutic tools for use against NSCLCs19,20,21. Furthermore, because the subcellular localization of RTKs regulate the downstream fate of RTK-elicited signals, the intracellular machineries involved in vesicle transport also represent potential targets of anti-cancer drugs15,22,23.

Several previous trials inferred a novel/hidden target of small-molecule compounds7,24,25,26. In this study, I developed a quantitative and statistical method to analyze microscopically obtained EGFR-related images. Fourteen inhibitors associated with signal transduction and intracellular trafficking of EGFR can be hierarchically classified based on their effects on cellular phenotype. I discovered that a 4,6-dianilinopyrimidine EGFR inhibitor (CAS 879127-07-8), the most uni-specific inhibitor among the various currently available kinase inhibitors27,28, was co-classified in the same cluster as the microtubule depolymerizer nocodazole. In fact, this compound induced microtubule depolymerization in both biochemical and cell-based assays. These data indicate that CAS 879127-07-8 could be used as a chemical probe to investigate the EGFR pathway and microtubule dynamics. The image-based multivariate analysis developed herein has potential as a powerful tool for discovering unanticipated drug properties.

Results

Quantitative analysis of signal transduction and intracellular traffic of EGF/EGFR

To examine the effects of various compounds on cellular phenotypes, I constructed an image-based assay system in which the intensity and intracellular localization of fluorescent signals were measured quantitatively. A549-GFP-EGFR cells, in which the genomic EGFR has been endogenously tagged with GFP, was used in this study. Cells were seeded in 96-well plates and treated for 1 h with inhibitors of EGFR signaling (Fig. 1A). EGF was then added to the culture at 100 ng/ml, a concentration at which EGFR was primarily transported to a degradation pathway29,30. After incubation for 0, 5, 30, 60, or 180 min, cells were fixed and processed for immunofluorescence using antibodies against molecules implicated in EGFR signaling, including phosphorylated ERK (pERK), phosphorylated Akt (pAkt) and several phosphoinositides (PtdIns(3)P, PtdIns(4)P and PtdIns(4,5)P2)22,31,32,33. In addition, endocytic trafficking was visualized using either EGF or transferrin. EGF was used as a marker to monitor the degradation pathway, whereas transferrin was used to measure the recycling pathway34,35,36. To visualize nuclear DNA, the cells were stained with Hoechst. Images were acquired by automated microscopy. Thus, cell phenotypes were monitored simultaneously using four different markers: GFP-EGFR, two signaling/trafficking molecules and Hoechst.

Figure 1
figure 1

Compounds and quantitative analyses.

(A) Inhibitors, components of the EGFR inhibitor panel (Merck Millipore) and nocodazole were used in this study. Primary targets of EGFR inhibitor panel are indicated in red. (B) Image segmentation and identification of objects to establish regions of interest (ROIs). DNA staining and GFP-EGFR signals were used to define the ‘nucleus’ and ‘cell’ regions, respectively. The ‘nucleus’ and ‘cell’ regions were expanded and shrunk, respectively and the resultant areas (after subtraction of the original objects) were defined as ‘perinuclear’ and ‘PM’. (C) An example of image segmentation. The input image of EGFR was segmented into the perinuclear (blue) and PM (green) regions and the signal intensity of GFP-EGFR (magenta) was measured in each region. (D) The ratios of EGFR signals in the perinuclear region to those in the PM region, measured in single cells treated with DMSO, were plotted for each incubation time.

Full size image

Several intracellular regions of interest were defined as illustrated in Fig. 1B. The ‘nucleus’ and ‘cell’ regions represent the area of Hoechst staining and GFP-EGFR fluorescence, respectively. The ‘perinuclear region’ was obtained by expanding the contour of the ‘nucleus’ with a diameter of 7 pixels, whereas the ‘plasma membrane’ region (PM) was obtained by shrinking the contour of the ‘cell’ region with a 5-pixel diameter. These intracellular regions were used to define the intracellular localization of observed signals/proteins37. In Fig. 1C, for example, the localization of EGFR was assigned to the ‘perinuclear’ and ‘PM’ regions. After the addition of EGF, I quantitatively monitored the trafficking of GFP-EGFR (Fig. 1D)35,38. The ratios of EGFR signal intensities from the ‘perinuclear’ region to those from the ‘PM’ region were measured in 500–1000 cells and plotted. As can be seen in the bar plot, EGFR moved toward the perinuclear region after incubation for 5 or 30 min (reflected as an increase in the number of cells with higher perinuclear signals). The regression coefficient (the slope of intensity in ‘perinuclear’ vs. ‘PM’) reached a maximum at 30 min, suggesting that EGFR reached to late endosomes/lysosomes, which localizes on perinuclear region (see Supplementary Fig. S1 as well). After incubation for 60 or 180 min, the number of cells with high signals decreased, indicating that EGFR was degraded. Thus, my selection of intracellular regions was appropriate for measuring the intracellular trafficking of EGFR.

Comparison of compounds’ effects using the Kolmogorov–Smirnov (KS) test

Using the system described above, I measured 151 image features (descriptors) listed in Supplementary Table S1. These features were selected comprehensively from the viewpoints of signal transduction and intracellular traffic. Figure 2A shows two examples of such descriptors (Nucleus_pERK and PM_pAkt). EGF stimulation induced phosphorylation of ERK and Akt and their translocation to the nucleus and PM, respectively. In control cells (DMSO), the signals from Nucleus_pERK and PM_pAkt reached their maxima after stimulation for 30 and 5 min, respectively. Treatment with certain compounds dramatically affected the appearance of these signals. For instance, inhibitors of EGFR (PD153035, CAS 879127-07-8 and PD168393) decreased both Nucleus_pERK and PM_pAkt. On the other hand, Akt Inhibitor VIII (an inhibitor of phosphorylation/activation of Akt) inhibited PM_pAkt but increased Nucleus_pERK, whereas PD98059 (an inhibitor of MEK, which phosphorylates ERK) inhibited Nucleus_pERK and increased PM_pAkt. These observations are consistent with the reported crosstalk between the Akt and ERK pathways39. Thus, quantitative image analysis employing descriptors is useful for revealing the effects of inhibitors.

Figure 2
figure 2

Statistical analysis of the effects of compounds based on their phenotypes.

(A) Integrated intensity of pERK from the ‘nucleus’ region (Nucleus_pERK) and pAkt from the ‘PM’ (PM_pAkt) were measured in each cell and kernel estimation was used to calculate the probability density estimation. (B) Key steps of algorithm for analyzing compound effects. A one-sided two-sample Kolmogorov–Smirnov (KS) test was performed to compare control and compound-treated cells and a standardized Z-score was calculated (see experimental methods). Descriptors with Z-scores exhibiting significant correlation between two independent experiments (R > 0.4) were used for subsequent analysis.

Full size image

To estimate the effects of inhibitors in a quantitative manner, I adopted the one-sided two-sample KS test (Fig. 2B, left panel)40,41. Because the conventional KS statistic lacks a sign, a variant statistic reported by Perlman et al. was used in this study24. A standardized Z-score was calculated by dividing the signed KS value with the standard deviation of the corresponding control KS value, which was calculated by bootstrapping simulation (n = 1000). This standardized/normalized Z-score was used to compare the extent of a compound’s effects on each descriptor.

The reliability of each descriptor was evaluated by calculating Pearson’s correlation coefficient between two independent experiments. If the correlation coefficient was above 0.4, descriptors were used for subsequent analysis; descriptors with lower coefficients were discarded. The correlation coefficients obtained for each descriptor are shown in Supplementary Table S1 and the selected descriptors are listed in Supplementary Table S2. Ultimately, 570 Z-scores were indexed by each of 114 descriptors at each of five time points (Zd, t). The Z-score was further indexed for each of 14 compounds (total, 7980 Zd, c, t) (Fig. 2B, right panel).

Cluster analysis of quantitative phenotypes categorized 14 compounds into 4 groups and CAS 879127-07-8 and nocodazole into the same cluster

Although the descriptors adopted in Supplementary Table S2 appear to represent many different types of information, it is possible that the data obtained for each descriptor may contain biological relevant content. Supplementary Fig. S2 depicts the correlation matrix detected among the Z-scores of 570 descriptors where the existence of positive or negative correlations can be recognized to various degrees. Hierarchical clustering of these correlations then shows seven major or fourteen minor groups (indicated by red and blue dotted lines, respectively, in Supplementary Fig. S3). The existence of such groups roughly suggests the extent of the data complexity. Therefore, I next tried to extract principal components (PCs) that would explain phenotype variability based on combinations of descriptors. Totally, fourteen PCs that could explain 100% of the variance were extracted and more than 95% of the variance could be explained by seven out of fourteen PCs (Supplementary Fig. S4). Considering both the number of observed molecules and their biological relevance, the detected numbers of groups such as seven or fourteen seems reasonable.

Figure 3A illustrates the relative contribution (score) of each descriptor (total, 570 descriptors) to each PC (four out of seven PCs are shown as representatives). PC1 was characteristic of highly negative scores for pAkt and EGF endocytosis, whereas positive scores for pERK made a large contribution to PC2. Furthermore, PC3 contained positive scores for pERK and negative scores for pAkt, suggesting that this component reflects, at least to some extent, the crosstalk between the ERK and Akt pathways39. PC4 contained highly positive scores for phosphoinositides, including PtdIns(3)P, PtdIns(4)P and PtdIns(4,5)P2 with some contribution from pAkt, suggesting that this component reflects phosphoinositide-linked pAkt activation42.

Figure 3
figure 3

(A) Contribution of descriptors to four major principal components. Descriptors are listed in Supplementary Table S1. (B) Hierarchical clustering of compounds based on seven principal components. Four main clusters were identified, as indicated by colors on the right side. (C) Heat map, based on data in Fig. 2A, shows the effects of inhibitors on PM transferrin intensity (left panel) and the PM endosome EGFR/transferrin count (right panel). (D) Cells were treated with the indicated compounds and allowed to internalize Alexa Fluor 647-EGF or Alexa Fluor 647-Tfn for 30 min.

Full size image

Using these seven PCs, I hierarchically clustered the observed effects of 14 inhibitors (and DMSO control) by the Euclid distance and Ward methods. As seen in Fig. 3B, four main clusters could be identified (Essentially similar results were obtained when using all of the fourteen PCs. See Supplementary Fig. S5). Because these four clusters were identified based solely on cellular phenotypes, it is possible that compounds falling into the same cluster might share common targets in signal transduction pathways.

One major characteristic of cluster 1 (and one not found in the other clusters) was a low PC1 score. In other words, in cluster 1 (but not in the other clusters), pAkt and EGF endocytosis was highly active. Therefore, compounds in cluster 1, including the DMSO control, are considered to exert relatively limited effects on signal transduction and trafficking of EGF/EGFR. These compounds include PP2 (Src inhibitor), AG490 (JAK inhibitor) and rapamycin (mTOR and S6K inhibitor), JNK inhibitor II, ZM336372 and PD98059. It should be noted that PD98059, an inhibitor of MEK, had an extremely low PC1 score; this compound might have stimulated massive phosphorylation Akt via cross-inhibition39.

On the other hand, compounds with high PC1 scores were divided among three clusters. Cluster 2 includes SB203580 (p38 MAPK inhibitor), Et-18-OCH3 (PI-PLC inhibitor), LY294002 (PI3K inhibitor) and Akt Inhibitor VIII. Several compounds affecting pathways upstream of Akt43,44,45,46,47 are included in this cluster. Cluster 3 contains PD153035 and PD168393; both are inhibitors of EGFR itself48,49. However, a third inhibitor of EGFR, CAS 879127-07-827, fell into cluster 4 along with nocodazole, an inhibitor of microtubule polymerization50, which is reflected to some extent by the high correlation coefficient of the Z-scores between these two compounds (See Supplementary Fig. S6). Because the presence of cluster 4 was totally unexpected, I sought to determine why CAS 879127-07-8 and nocodazole were in the same cluster.

Nocodazole does not likely target kinases involved in the EGFR pathway

Although CAS 879127-07-8 is a highly uni-specific inhibitor28, many other kinase inhibitors have one or more secondary targets in addition to their primary target (for an illustration, see CAS 879127-07-8, gefitinib and erlotinib in Supplementary Fig. S7A, which depicts a drug profiling panel using recombinant kinases4). Thus, I explored the possibility that nocodazole might target a kinase as well as microtubules. The target molecules reported to date for the compounds in clusters 3 and 4 are provided in Supplementary Fig. S7B (data from ChEMBL). Three inhibitors of EGFR share two targets, EGFR and ErbB4, whereas nocodazole has no kinase target. Target prediction using SwissTargetPrediction also indicated that nocodazole does not share any kinase target with CAS 879127-07-8 (Supplementary Fig. S7C). Although one recent study reported that nocodazole is a high-affinity ligand for Abl, c-kit, B-Raf and MEK51, I observed hyperphosphorylation of ERK in nocodazole-treated cells as described by Hayne et al., suggesting that the Raf-MEK-ERK pathway is not likely to be impaired by nocodazole. Thus, it seemed unlikely that CAS 879127-07-8 and nocodazole target a common kinase/pathway.

CAS 879127-07-8 and nocodazole induce similar cellular phenotypes

PC analyses were adequate to classify inhibitors, as shown in Fig. 3B. However, because each PC was actually a complicated mixture of descriptors, such analysis was not sufficient to characterize each inhibitor in detail. Therefore, I searched one by one for descriptors shared by nocodazole and CAS 879127-07-8 (but not by the other inhibitors of EGFR, PD153035 and PD168393). Two descriptors satisfied these requirements: the intensity of transferrin signals around PM (PM transferrin intensity) and the count of PM endosomes containing EGFR/transferrin (PM endosome EGFR/transferrin count).

As seen in Fig. 3C, both CAS 879127-07-8 and nocodazole increased PM transferrin intensity and the PM endosome EGFR/Tfn count (Fig. 3C and red arrowheads in Fig. 2B, right panel). After internalization from the PM, transferrin is delivered to early endosomes and subsequently to recycling endosomes. Therefore, the observed increase in PM transferrin intensity indicates that internalization of transferrin from the PM was impaired. On the other hand, EGF/EGFR internalized into early endosomes is directed toward a degradation pathway. The observed increase in the PM endosome EGFR/transferrin count means that the sorting of EGFR and transferrin at early endosomes was impaired35,36 and that these two molecules coexisted in that compartment for a prolonged period of time.

Figure 3D shows a microscopic image of transferrin/EGF behavior. Because transferrin internalization occurs independently of EGFR phosphorylation, PD153035 (and PD168393, not shown) did not affect transferrin internalization (Fig. 3D, bottom panel). By contrast, in cells treated with CAS 879127-07-8 or nocodazole, transferrin accumulated at the PM even after incubation for 30 min. On the other hand, all three compounds (PD153035, CAS 879127-07-8 and nocodazole) impaired EGF internalization (Fig. 3D, upper panels). Both PD153035 and CAS 879127-07-8 inhibit EGFR auto-phosphorylation, a step required for its internalization52. By contrast, nocodazole disrupts microtubule organization required for the intracellular transport of internalized vesicles, causing EGF/transferrin to accumulate in the PM/endosome region. Thus, the inhibition of transferrin internalization observed upon treatment with CAS 879127-07-8 might be due not to this compound’s inhibitory effect on EGFR, but instead to a mechanism similar to that of nocodazole.

CAS 879127-07-8 is a novel microtubule depolymerizer

Based on the above reasoning, I analyzed the possible effects of CAS 879127-07-8 on microtubule organization (Fig. 4A). In control and PD153035-treated cells, I observed a radial array of microtubules. On the other hand, microtubule organization was disrupted in CAS 879127-07-8- and nocodazole-treated cells. Similar results were obtained in seven other cell lines of different genetic backgrounds (Supplementary Fig. S8). One of these cell lines, NCI-H661 does not express endogenous EGFR53. Therefore, the observed disruption of microtubules by CAS 879127-07-8 appears to be unrelated to its inhibitory effect on EGFR tyrosine kinase activity. In addition, the addition of CAS 879127-07-8 or nocodazole, but not PD153035, to the culture reduced the cell number and increased the percentage of cells in G2/M phase, probably because mitosis was impaired (Supplementary Fig. S9). Thus, like nocodazole, CAS 879127-07-8 disrupts cellular microtubule organization, as nocodazole does.

Figure 4
figure 4

CAS 879127-07-8 is a novel microtubule depolymerizer.

(A) Cells were treated with indicated compounds for 1 h, fixed and processed for immunofluorescence using anti-tubulin antibody. Polymerized microtubules appeared in DMSO- and PD153035-tretaed cells, but not in CAS 879127-07-8- or nocodazole-treated cells. (B) In vitro microtubule polymerization assay confirmed that both CAS 879127-07-8 and nocodazole, but not PD153035, prevented microtubule polymerization in a concentration-dependent manner (center panel). (C) Cells were counted after treatment with the indicated compounds for 24 h. Note that CAS 879127-07-8 and nocodazole, but not PD153035, decreased cell number, indicating that these compounds interfered with mitosis.

Full size image

Next, I investigated whether CAS 879127-07-8 disrupts microtubule organization through a direct or indirect action on microtubules. In an in vitro microtubule polymerization assay, CAS 879127-07-8 and nocodazole (but not PD153035) inhibited tubulin polymerization in a concentration-dependent manner (Fig. 4B), indicating that CAS 879127-07-8 acts directly on microtubule polymerization. Collectively, these data show that CAS 879127-07-8 has a dual activity: microtubule depolymerization and kinase inhibition.

Structural insight of dual activity of CAS 879127-07-8

Finally, I investigated the structural basis of the dual activity of CAS 879127-07-8. Models of the complexes between CAS 879127-07-8 and EGFR and between CAS 879127-07-8 and tubulin, were constructed by docking simulation (Fig. 4C). The simulation predicted that CAS 879127-07-8 docks into an ATP-binding site of EGFR. This prediction is in accordance with a report27 in which binding of CAS 879127-07-8 to the ATP-binding site was confirmed biochemically. Notably, for CAS 879127-07-8 to fit into the ATP-binding site, it must adopt an extended structure (Fig. 4C). On the contrary, it must assume a bent configuration to dock in the interface between alpha- and beta-tubulin. CAS 879127-07-8 has seven rotatable arms, which might contribute to its structural flexibility between the extended and bent forms.

Discussion

In this study, I showed that CAS 879127-07-8, a previously described EGFR inhibitor, also affects microtubule polymerization. Thus, this compound affects two unrelated molecules, both of which are targets of anti-cancer drugs. Mutations of EGFR are frequently observed in various cancers, whereas inhibitors of microtubules are capable of blocking mitosis. As a blocker of mitosis, CAS 879127-07-8 suppressed cell growth in a concentration-dependent manner (Supplementary Fig. S9).

I evaluated the phenotypic effects on EGFR signaling and trafficking using 14 commercially available compounds related to the EGFR signaling pathway. To this end, I subjected microscopic images to high-content screening using automated microscopy and image analysis software. For the image analysis, I defined two intracellular regions, perinuclear and PM, which worked well as reported previously37. Subsequently, image analyses were combined using classical statistical methods, such as the KS test, PC analysis and hierarchical clustering.

As a result of these combined analyses, the 14 compounds were classified into four clusters. The separation of cluster 1 from the other three reflected differences in the PC1 scores, which were heavily influenced by phosphorylated Akt. Half of the compounds examined, including the DMSO control, were classified into the PC1. However, this does not necessarily mean that these compounds have small effects on EGFR signaling. In this study, I selected two major signaling molecules, ERK and Akt, which represent the Ras-MEK-MAPK pathway and PtdIns3K-Akt-mTOR pathway, respectively. Other signaling pathways, including the p38MAPK, PKC, JAK-STAT and JNK pathways, are also important for cell growth and tumorigenesis. Adoption of these signaling molecules for phenotypic observation in the high-content screen might separate some of the compounds in cluster 1 from the DMSO control.

One major finding in this study was that CAS 879127-07-8 was categorized into cluster 4 together with nocodazole. This was totally unexpected because CAS 879127-07-8 was reported to be a uni-specific inhibitor with only one target among 300 kinases tested. In its extended form, CAS 879127-07-8 binds to an ATP-binding pocket as gefitinib does27. The chemical structure of CAS 879127-07-8 consists of three benzenes linked by seven rotatable arms, which may endow CAS 879127-07-8 with structural flexibility. My simulation study predicted that CAS 879127-07-8 binds to the interface between alpha- and beta-tubulin similarly to the manner predicted in silico for the microtubule depolymerizer colchicine54. Thus, CAS 879127-07-8 could be used as a seed to develop a novel anti-cancer drug that simultaneously inhibits two unrelated proteins, EGFR and microtubules. Because EGFR is one of the major driving forces in cancer metastasis, it is a promising candidate of anti-cancer drugs; consequently, gefitinib and erlotinib are widely used in the treatment of lung cancer16,17,18. Microtubule inhibitors, some of which are used as anti-cancer drugs, can inhibit mitosis55. This study showed that CAS 879127-07-8 also inhibits mitosis, as shown in Supplementary Fig. S9. Thus, CAS 879127-07-8 functions as a uni-specific kinase inhibitor and a microtubule depolymerizer. One other molecule, S9, is known to be a dual inhibitor of microtubules and a kinase; such dual-target drugs represent a new strategy for the targeted therapy of cancer56.

In summary, I developed a quantitative and statistical system for analyzing microscopic images that reflect the effects of various inhibitors on signal transduction and vesicle trafficking of EGFR. I demonstrated that a highly specific kinase inhibitor, CAS 879127-07-8, also functions as an inhibitor of microtubules as well. Thus, this system appears to be a powerful tool for discovering unexpected properties of known inhibitors. In the future, this system should be applicable to studies of other RTKs, as well as the relationship between intracellular signaling and vesicle trafficking.

Methods

Cells and reagents

A549 GFP-EGFR cells, which express GFP-tagged endogenous EGFR protein, were purchased from Sigma (CLL1141) and maintained at 37 °C in DMEM containing 10% FBS and antibiotic-antimycotic solution (A5955, Sigma). A549 cells and the lung cancer cell panel were purchased from the RIKEN BioResource Center and ATCC, respectively and maintained in RPMI-1640 containing 10% FBS (except for H1573, whose medium contains 5% FBS) and antibiotic-antimycotic solution. Recombinant EGF and human transferrin were purchased from R&D Systems. Alexa Fluor 647-conjugated EGF, Alexa Fluor 647-conjugated human transferrin and Hoechst 33342 were from Invitrogen. Rabbit anti-phosphorylated ERK monoclonal antibody and rabbit anti-phosphorylated Akt monoclonal antibody were purchased from Cell Signaling Technologies. Mouse anti-PtdIns(4)P monoclonal antibody and mouse anti-PtdIns(4,5)P2 monoclonal antibody were purchased from Echelon. Rabbit anti-GST polyclonal antibody was purchased from Santa Cruz Biotechnology. Alexa Fluor-conjugated secondary antibodies were purchased from Invitrogen. All antibodies were preserved at −30 °C after addition of an equal volume of glycerol. Compounds were purchased as an EGFR inhibitor panel (Merck Millipore) except for nocodazole (Sigma).

Recombinant proteins

pGEX-HrsFYVE was transformed into BL21 (DE3) and recombinant protein was induced by incubation with 0.1 mM IPTG for 1 h at 37 °C. Cells were lysed for 30 min at 4 °C using B-PER bacterial protein extraction reagent (Pierce) containing CompleteTM protease inhibitor cocktail (Roche). Cells were clarified by centrifugation (1 min at 15,000 × g) and the resultant supernatants were rocked with glutathione-Sepharose 4B (Amersham) for 1 h. Beads were collected by centrifugation (5 min at 600 × g) and then washed three times with B-PER. GST-HrsFYVE was eluted from the beads with B-PER containing 20 mM glutathione. Eluted proteins were frozen and used for experiments.

Inhibitor treatment and EGF stimulation

Cells were seeded on Edge-plates (Thermo Scientific), serum-starved for 6 h by replacing the medium with DMEM containing 0.1% BSA and stimulated with 100 ng/ml EGF (R&D Systems). One hour before EGF stimulation, cells were treated with various compounds at the indicated concentrations: Akt Inhibitor VIII (200 nM)57, PD153035 (100 nM)58, CAS 879127-07-8 (10 μM)27, Et-18-OCH3 (5 μg/ml)59, JNK inhibitor II (1 μM)60, LY294002 (10 μM)57, PD98059 (50 μM)61, PD168393 (100 nM)62, PP2 (100 nM)63, rapamycin (10 nM)64, SB253080 (10 μM)60, AG490 (40 μM)65, ZM336372 (1 μM)66 and nocodazole (50 μM)67. To observe internalized EGF, Alexa Fluor 647-conjugated EGF was used for initial internalization and chased after 5 min with unlabeled EGF. To observe internalized transferrin, Alexa Fluor 647-conjugated transferrin and unlabeled EGF were internalized for 5 min and then chased with unlabeled human transferrin (R&D Systems) and EGF. After the indicated intervals, cells were fixed by addition of an equal volume of 4% paraformaldehyde (09154-85, Nacalai Tesque) for 15 min, followed by washing in PBS.

Immunofluorescence

Fixed cells were subsequently processed for immunostaining. To stain phosphorylated ERK, aldehyde-fixed cells were further incubated with ice-cold methanol for 10 min at −30 °C, washed with PBS and incubated in blocking buffer (1% BSA in PBS) containing 0.3% Triton X-100 (TX100) for 1 h at RT. To stain phosphorylated Akt, fixed cells were directly incubated in blocking buffer containing TX100. To stain phosphoinositides, fixed cells were permeabilized with 0.01% digitonin in PBS for 5 min at RT, washed with PBS and incubated in blocking buffer without TX100 for 1 h at RT. To observe PtdIns(3)P, recombinant GST-Hrs FYVE protein was added to the blocking buffer and after incubation for 1 h at RT, cells were washed with PBS. Blocked cells were incubated with the following primary antibodies for 1 h at RT: anti-pERK (1:1000), anti-pAkt (1:2000), anti-PtdIns(4)P (1:2000), anti-PtdIns(4,5)P2 (1:1000), or anti-GST (1:200). Cells were washed with PBS two times and subsequently incubated for 1 h at RT with the appropriate secondary antibodies and Hoechst 33432. Cells were washed twice with PBS, fixed with 2% paraformaldehyde for 5 min at RT, washed with PBS and stored at 4 °C.

Image acquisition and analysis

Stained cells were photographed using a cell image analyzer (CellInsight, Thermo Scientific) equipped with a 20× objective lens. Hoechst and GFP signals were used in all marker sets to define the ‘nucleus’ and ‘cell’ regions, respectively. In each well, 36 fields were photographed. Image analysis was performed using CellProfiler (Broad Institute) and all images were processed for illumination correction. Using the primary object ‘nucleus’ and GFP-EGFR signals, the cell outline (‘cell’) was identified by the ‘propagation’ implemented in CellProfiler. ‘Cytoplasm’ was defined by the subtraction of ‘nucleus’ from ‘cell’. The ‘nucleus’ and ‘cell’ regions were expanded and shrunk by 5 and 7 pixels, respectively and the resultant regions were named ‘perinuclear’ and ‘PM’, respectively. These regions of interest (ROIs) were used to quantitate each image feature. The signals of some fluorescent signals were converted into objects to identify ‘endosomes’ using the Top-Hat filter implemented in the ‘enhance’ module in CellProfiler. Co-localization was also evaluated using these objects. Descriptors used in this study are listed in Supplementary Table S1.

Data processing and statistical analysis

Data was stored in a local MySQL server and processed for statistical analysis using MatLab (Mathworks). Selected image features (descriptors) of each compound were compared with control (DMSO-treated) cells using the one-sided two-sample KS test. Addition of a sign to KS statistics was performed according to the method of Perlman et al.24. Signed KS statistics were divided by the corresponding control standard deviation (Std), which was estimated by bootstrap simulation with replacement (n = 1000) to yield a standardized Z-score. When the correlation coefficient of the Z-score between two independent experiments was above 0.4, the descriptor was accepted for subsequent analysis. A total of 570 Z-scores were accepted and then subjected to principal component analysis (PCA). Seven PCs explained over 95% of the variance and were used for subsequent hierarchical clustering using the Euclid distance and Ward method.

Bioinformatics

Known targets of compounds were compiled using information from the ChEMBL database (EMBL-EBI) and selected according to the following criteria: 1) the IC50 is below the concentration used in this study; 2) the target is indicated as a ‘single protein’; and 3) the entry was curated by an ‘expert’ (not ‘autocuration’). Target prediction was performed using SwissTargetPrediction (SWISS-SBI). Docking simulation was performed using CLC Drug Discovery Workbench 2.1 (Qiagen).

Additional Information

How to cite this article: Tanabe, K. Image-based compound profiling reveals a dual inhibitor of tyrosine kinase and microtubule polymerization. Sci. Rep. 6, 25095; doi: 10.1038/srep25095 (2016).