Despite the development of a number of efficacious kinase inhibitors, the strategies for rational design of these compounds have been limited by target promiscuity. In an effort to better understand the nature of kinase inhibition across the kinome, especially as it relates to off-target effects, we screened a well-defined collection of kinase inhibitors using biochemical assays for inhibitory activity against 234 active human kinases and kinase complexes, representing all branches of the kinome tree. For our study we employed 158 small molecules initially identified in the literature as potent and specific inhibitors of kinases important as therapeutic targets and/or signal transduction regulators. Hierarchical clustering of these benchmark kinase inhibitors on the basis of their kinome activity profiles illustrates how they relate to chemical structure similarities and provides new insights into inhibitor specificity and potential applications for probing new targets. Using this broad dataset, we provide a framework for assessing polypharmacology. We not only discover likely off-target inhibitor activities and recommend specific inhibitors for existing targets, but also identify potential new uses for known small molecules.
- DFG motif
- functional genomics
The reversible phosphorylation of proteins and lipids is one of the central themes within cellular growth and differentiation signalling pathways. Since protein kinases play a fundamental role in cell signalling, especially as components of critical growth and differentiation pathways linked to cancer, kinase inhibitors have found applications as both therapeutics and research tools. This is due, in part, to the chemical tractability of protein kinase ATP-binding pockets, the automation of activity measurements, and the correlation of in vitro, cellular and in vivo activities. The 518 human kinases that are responsible for transferring phosphate on to regulatory serine, threonine and tyrosine residues on proteins have been successfully exploited as both drug and research probe targets [1–5]. The success of kinase inhibitors as drugs, as exemplified by the small-molecule tyrosine kinase inhibitor, imatinib (Gleevec), has focused even greater emphasis on development of this class of chemical entities as potential breakthrough therapeutics [6–8]. Similar to Gleevec, the majority of kinase inhibitors developed thus far target ATP-binding sites; this suggests that related kinases with similar active sites may represent potentially problematic unintended targets. Initial studies related to deleterious off-target activities have focused on closely related proteins defined by sequence or structural similarity [9–11]. With the advent of newer profiling technologies that allow higher-throughput analysis of large numbers of targets and target families, we can now include ‘activity fingerprint similarity’ as an additional strategy to screen for off-target effects . Specificity issues are critical for understanding both detrimental off-target effects and for defining an improved rational drug development strategy on the basis of polypharmacology [13,14].
Historically, both drug and chemical probe development strategies have been driven using a target-centric approach. That is, the initial goal was to identify potent molecules against a single target and then test off-target effects using a small number of candidate compounds. The proliferation of profiling technologies now allows early selectivity and potency profiling against a more diverse set of compounds in order to quickly assess the first principles and guidelines of chemical probes as defined by Frye  and Cohen . That is, a good chemical probe has sufficient in vitro potency and selectivity data to confidently associate to its cellular in vivo profile. Fulfilling this first principle has proven difficult, but is becoming more straightforward as more efficient screening tools become available. This is relevant not only for the development of small molecules as research tools, but also can be applied when defining potential therapeutics.
The need for specific inhibitors to dissect signalling pathways and validate targets is especially important for the biological research community. There is a lack of comprehensive data available to explain biological findings, especially those related to the effects of compounds used in high content analysis and phenotypic screens. This lack of data is hindering our understanding of biological observations and in some cases is likely to lead to incorrect conclusions.
A number of kinase profiling platforms that allow screening across the complete kinome have been advanced as economical tools for addressing the specificity issue. These have been used to generate information resources for the scientific community, especially as it relates to structure–function relationships for kinase inhibitor development [17–19]. For example, Davis et al.  have defined selectivity profiles for 72 inhibitors using a robust high-throughput binding assay that relied on competition of biotin-labelled staurosporine to calculate Kd for bacterially expressed T7 bacteriophage human kinase fusions. The binding data generated using this platform has generally compared favourably with standard radioactive assays that measure the biochemical phosphotransferase activity of the kinase. A second technology proposed for kinome profiling involves a thermal stability shift assay . This approach uses changes in protein stability in the presence of stabilizing inhibitors to identify molecules with sub-micromolar IC50 values on the basis of correlations with activity assays. Although there is often reasonably good correlation between these binding assays and traditional activity screens, profiling on the basis of binding alone exhibits significant false-positive and false-negative rates for predicting inhibition of catalytic activity . Additionally, newer chemoproteomics kinase profiling platforms utilizing probes for accessing small-molecule selectivity for multiple kinases in cell extracts have been utilized [21–23]. For example, Patricelli et al.  identified differences in potency and selectivity for inhibitors when comparing this method with standard profiling of recombinant kinases in vitro.
To define the relationship between kinase targets and chemistry further, we profiled 158 structurally diverse small molecules, originally defined in the literature as potent and specific kinase inhibitors, for their inhibitory activity towards 234 human recombinant kinases. We employed the well-accepted standard in vitro phosphotransferase activity assay using specific peptide substrates or proteins and ATP concentrations near Km in order to ensure equivalence. Assays were performed at two inhibitor concentrations, 1 μM and 10 μM, ensuring internal control as well as an estimation of potency. The results of the present study, including more than 70000 data points, are consistent with previous profiling screens [19,24–26] and significantly add to the findings of these studies. When combined with traditional chemical structure similarity analysis, hierarchical clustering of the inhibitors on the basis of kinome activity profiles can generate new insights into inhibitor specificity and applications for probing new targets. We not only observe differences in activity profiles for structurally similar inhibitors, but also identify new targets for previously characterized tool compounds. These results have implications for the development of strategies based on polypharmacology and predicting off-target effects. Our results can be used as a guide to allow researchers to pick the most relevant research tool compounds for their specific applications, to better understand research findings when these compounds are used, and to provide additional context for therapeutic target selection and drug discovery strategies.
A total of 234 purified human recombinant kinases and kinase complexes were expressed as either full-length proteins or catalytically active fragments (Supplementary Table S1 at http://www.biochemj.org/bj/451/bj4510313add.htm). Kinase assays were performed as per the EMD Millipore profiling service protocol (http://www.millipore.com/techpublications/tech1/pf3036) using peptide or protein substrates in a filter-binding radioactive ATP transferase assay. Activity assays in the absence of inhibitors (DMSO control) were performed in quadruplicate and assays performed in the presence of control inhibitors (Supplementary Table S1) were performed in duplicate. Inhibitor control wells contain all components of the reaction including a previously characterized reference inhibitor. For kinases where no suitable inhibitor exists, 30% phosphoric acid is used in place of the inhibitor control. Following subtraction of the average inhibitor control well counts, results were expressed as a percentage of the mean kinase activity in the positive control sample. Assays performed in the presence of 158 different kinase inhibitors (Calbiochem Protein Kinase Inhibitor Library I and II, catalogue numbers 539744 and 539745; Supplementary Table S2 at http://www.biochemj.org/bj/451/bj4510313add.htm) were performed in singlicate at two concentrations of 1 μM and 10 μM. Such an approach considers two different inhibitor concentrations that can be leveraged as internal controls (quasi-replicates), but also establishes an estimated dose–response relationship. IC50 values for select inhibitors were calculated from a nine-point dose–response curve generated using the standard assay conditions for the target as described above. The ten-compound validation screen was performed exactly as above except in duplicate. Standard errors were calculated on the basis of percentage remaining activity.
Km values for ATP were determined for each kinase using standard assays as described above over an appropriate range of ATP concentrations. The Km value for ATP was calculated using standard velocity compared with substrate (ATP) concentration assuming first-order kinetics.
Potency and selectivity scoring and clustering
A potency value was assigned for each compound–kinase pair on the basis of the capacity of the compound to reduce effective activity relative to a DMSO-treated control. Each compound was assigned one of four potency categories for each kinase target: inactive (0), weakly active (1), active (2) or very active (3) (see Figure 2A). The following criteria were used to assign potency levels for each inhibitor towards each of the 234 individual kinases. Inhibitors with remaining activity ≤10% at both 1 μM and 10 μM were assigned an activity level of 3 and represent very active compounds for that target (IC50 ≪1 μM). Inhibitor–kinase pairs, not assigned as above, with remaining activity ≤10% at 1 μM or 10 μM were assigned an activity level of 2 and represent active compounds for that target (IC50 ≪10 μM). Remaining uncategorized inhibitor–kinase pairs with remaining activity <50% at 1 μM or 10 μM were assigned an activity level of 1 and represent a weakly active compound for that target (IC50 ~10 μM). Inhibitor–kinase pairs with remaining activity ≥50% at 1 μM and 10 μM were assigned an activity level of 0 and represent inactive compounds for that particular kinase. Inhibitor potency was visualized using a radar plot rendered in Excel. Potency is indicated by the length of the vector (0–3) for each kinase with an origin in the centre of the dartboard. The kinase targets themselves are distributed around the perimeter of the dartboard in a clockwise manner from the most sensitive kinase in general (that is, the kinase inhibited by the greatest number of small molecules) to the least sensitive kinase (as described in the Kinase target sensitivity section).
Inhibitor selectivity scores (S) and kinase sensitivity scores (S’) represent fractions of activities among subjects screened. That is, the S score for an inhibitor equals the number of kinases inhibited with a potency of 2 or 3 divided by the total number of kinases screened (234). The S’ score for a kinase equals the total number of inhibitors that inhibit the kinase with a potency of 2 or 3 divided by the total number of inhibitors profiled (n=58).
Kinase hierarchical clustering analysis was executed using Ward's method, contained within the standard statistical package of R. The distance matrix was calculated using the R binary method. Activity levels of 2 or 3 (see the Potency assignment section below) in this screen are treated as active and activity levels of 0 and 1 are treated as inactive. The distance of object pairs is the proportion of elements in which only one object is active amongst those in which both objects are active. The binary method applied in the present study reveals compounds that are similar as a result of being active towards similar kinases. The clustering results were visualized as phylogenetic trees using the software package Mega5 (Molecular Evolutionary Genetic Analysis, http://www.megasoftware.net/).
Small molecule structure similarity analysis and hierarchy clustering was performed using the PubChem Chemical Structure Clustering Tool (http://pubchem.ncbi.nlm.nih.gov/assay/?p=clustering) available in the public domain and based on PubChem 2D (two-dimensional) structure fingerprints (ftp://ftp.ncbi.nlm.nih.gov/pubchem/specifications/pubchem_fingerprints.txt). Chemical structures can be described as PubChem binary substructure fingerprints, which are ordered lists of binary (1/0) bits. Each bit represents a Boolean determination of the presence of, among others, an element count, a type of ring system, atom pairing or atom environment (nearest neighbours) in a chemical structure. Compound pair similarities were then calculated using a Tanimoto similarity measure.
We assayed the activity of 158 structurally diverse small molecules against 234 human recombinant kinases. The kinases screened are depicted in Figure 1 and have a significant degree of overlap with the targets assayed in a recent publication on kinase profiling . The kinases and published kinase inhibitor targets are distributed evenly across the human kinome, with representative targets from all major branches of the tree, and include most kinase drug targets currently identified. Lipid kinases were not analysed so that we could employ a consistent detection method using a single platform across the screen. Figure 1 depicts coverage of the kinome tree with approximately 45% of the kinome (red dots) screened. The 158 kinase inhibitors, most of which are reversible ATP-competitive Type I inhibitors , were initially chosen on the basis of literature publications that defined their targets as important signal transduction and disease-associated kinases. Of the kinases profiled, 34% are targeted by at least one of the 158 inhibitors based on publications. Additionally, the collection of inhibitors was selected to be structurally diverse and stable in solution (Supplementary Table S2).
In almost all cases the assays were carried out under reaction conditions that included ATP concentrations within 3-fold of the measured Km value for each kinase. This is different from most screens, especially those based on binding assays, where the ATP concentration is generally at a single fixed concentration or is omitted. Since a majority of the inhibitors screened are ATP-competitive, using a fixed ATP concentration may affect the relative rank order of sensitivity to an inhibitor for those targets with a very high Km value for ATP. Specific peptides or protein substrates, buffers and ATP concentrations were optimized for each assay individually (Supplementary Table S1). The results of the assays are shown as a percentage of the remaining activity relative to activity in the absence of inhibitor (DMSO control) (Supplementary Table S3 at http://www.biochemj.org/bj/451/bj4510313add.htm). Background signal was normalized on the basis of previously characterized standard control inhibitors used as the baseline (Supplementary Table S1). As described in the Experimental section, all compounds were assayed at 1 μM and 10 μM. Inhibitor concentrations of 1 μM and 10 μM were chosen since these are the historical standard concentrations for primary screens as well as the fact that effective concentrations of these inhibitors tend to be higher in vivo than in vitro. This ensured detection of any off-target effects that may result from the use of higher concentrations of inhibitors normally used when interrogating pathways within cells .
To assess assay reproducibility in an efficient manner, we performed an additional validation screen of ten representative small molecules contained in the initial 158 molecule collection (Supplementary Table S4 at http://www.biochemj.org/bj/451/bj4510313add.htm). The additional assays were performed in duplicate at both 1 μM and 10 μM exactly as the initial large screen. The representative molecules chosen for use in this second screen reflect the broad spectrum of selectivity attributed to the compounds used in the initial screen, ranging from very selective (S=0.01) to promiscuous (S=0.41) as defined by their S scores (Table 1, and see the Selectivity of inhibitors section below). The average standard error for all 234 kinase assays across ten inhibitors in duplicate was 2.95 (range=1.10–8.25) and showed no apparent bias related to inhibitor selectivity or concentration (Table 1). Additionally, when including the single assay point from the initial larger screen in the statistical calculations, the average standard error was increased to 5.85 (range=2.34–13.03), most likely representing inter-assay variability related to different kinase preparations, inhibitor preparation and experimental days. The statistical analysis of error attests to the reliability of the assays and data for determination of potency and selectivity of the inhibitor collection across all 234 kinases.
Potency was assigned for each inhibitor–kinase pair as described in the Experimental section and in Figure 2(A). As an example, representative potency assignments for the Src family kinase inhibitor SU6656 [CAS (Chemical Abstract Service) 330161-87-0] across 30 kinase targets are depicted in Figure 2(B). Assignment of such potency scores enables a rough estimation of compound potency, even though we have not performed multiple point dose–response assays for all inhibitor–kinase pairs. This served as the starting point for all our further data analysis. Using this assignment we distributed the data across the potency spectrum with 1530 data points designated as very active, 1943 data points as active, 4804 data points as weakly active and 29163 data points as inactive (Supplementary Table S5 at http://www.biochemj.org/bj/451/bj4510313add.htm).
To validate our potency assignment, two inhibitor–kinase pairs were further characterized with follow-up IC50 determination using a nine-point dose–response analysis. SU6656 (CAS 330161-87-0), initially assigned a potency of 3 (very active) against Lck (Figure 2B), displayed an IC50 of 77 nM (Figure 2C), confirming the correct potency assignment (IC50 ≪1 μM). GSK (glycogen synthase kinase) 3β Inhibitor VIII (CAS 487021-52-3) was initially assigned a potency of 2 (active) for GSK3β and displayed an IC50 of 147 nM in the follow-up assay (Figure 2C). Additional potency analyses of Akt inhibitor X (CAS 925681-41-0) and the kinase PASK (Per/Arnt/Sim domain-containing serine/threonine-protein kinase) will be discussed later.
We have compared the data with a recent publication that analysed a similar set of kinase inhibitors in order to determine consistency and identify any discrepancies resulting from differences in assay platforms or data processing. The overlap between kinases and inhibitors tested in the present study and in the recent publication by Anastassiadis et al.  is significant; however, we have evaluated 25 kinases not tested in that study, including important drug discovery and signalling targets such as mTOR (mammalian target of rapamycin), AMPK (AMP-activated protein kinase) and eEF2K (eukaryotic elongation factor 2 kinase), as well as the cyclin/CDK (cyclin-dependent kinase) complexes (Supplementary Table S6 at http://www.biochemj.org/bj/451/bj4510313add.htm). Although the data generated by Anastassiadis et al.  used a similar profiling platform based on traditional radioactive filter binding activity assays, there are some important differences. First, in almost all cases we used ATP concentrations within 3-fold of the Km value for each kinase (11 outlier kinase concentrations were within 11-fold of the measured Km value for ATP), whereas Anastassiadis et al.  used a fixed ATP concentration of 10 μM. Secondly, the screening of two inhibitor concentrations allows us to determine a concentration-dependent inhibition effect which should be more accurate than using a single concentration of inhibitor (0.5 μM) as employed by Anastassiadis et al. . Use of a single lower concentration may not allow detection of weakly active compounds that may have off-target effects when used at relatively high doses in cells or organisms, especially for those inhibitors that hit kinases with differing Km values for ATP . Nevertheless, the data are generally consistent with Anastassiadis et al. . Among the shared inhibitor–kinase pairs assigned an activity level of 3 (very active), 84% are reported to display greater than 50% inhibition by Anastassiadis et al.  (Supplementary Figure S1 at http://www.biochemj.org/bj/451/bj4510313add.htm). Additionally, 60% of the shared inhibitor–kinase pairs assigned an activity level of 2 (active) or 3 (very active) are reported to have greater than 50% inhibition in the study by Anastassiadis et al. . When comparing the results from the other direction, 74% of the shared inhibitor–kinase pairs with greater than 50% inhibition as determined by Anastassiadis et al.  were assigned an activity level of 2 (active) or 3 (very active). Furthermore, readily apparent dose dependency is demonstrated when combining results across the shared targets. For example, a snapshot of assays against Syk inhibitor II (CAS 227449-73-2) incorporating both sets of data at three different concentrations (0.5 μM from Anastassiadis et al.  and 1 μM and 10 μM from the present study) demonstrates that most compounds showed a consistent level of inhibition (Figure 3). However, there are examples where we have detected inhibition in our profile that was not identified by Anastassiadis et al. . For example, the previous study did not identify CaMK (Ca2+/calmodulin-dependent protein kinase) I or CaMKII as targets of the Syk inhibitors, but our results clearly indicate CaMKIIβ, CaMKIIγ and CaMKIIδ were all inhibited by Syk inhibitor II at both 1 μM and 10 μM, the doses typically used in cellular assays (Figure 3) [29,30].
Selectivity of inhibitors
Using the above assignments for potency, we determined a Selectivity Score (S) for each small-molecule inhibitor on the basis of the number of targets inhibited at high potency (potency score of 2 or 3). Assignment of selectivity scores can be useful when prioritizing potential leads for target-specific optimization or in identifying more appropriate tool compounds. We chose the S score, which is intuitive and robust for a large-scale kinome screen and is defined as the number of kinases inhibited (in our case a potency score of 2 or 3) divided by the total number of kinases in the profiling. This gives a score between 1 (most promiscuous) and 0 (most selective) and can be used as a general measure even when using smaller profiling sizes . In the present study we determined that the median number of kinase targets potently inhibited by a small molecule is six (Figure 4A). At one end of the spectrum, 194 of the 234 kinase targets were inhibited by staurosporine (S=0.83), the least selective compound screened. At the other end of the spectrum, only a single target was hit by JNK (c-Jun N-terminal kinase) inhibitor IX (CAS 312917-14-9), making it the most selective inhibitor (S=0.004) in the present study (Figure 4A and Supplementary Table S5). A total of 25 small molecules displayed little or no inhibition (potency score of 0 or 1), presumably because the intended target initially identified in the literature was not included in the profile (i.e. lipid kinases) or the in vitro assay employed was different from the original assay reported in the literature either in platform or actual target character (i.e. full-length protein compared with fragment or mutant kinase).
Similar to other published large profiling screens, we detected selective and non-selective compounds targeting all branches of the kinome tree as well as inhibition of distantly related kinase targets by the same molecule. For example, it is clear that the TGFβR (transforming growth factor-β receptor) I inhibitors are very selective for their originally published targets (ALK4 and ALK5) and that Aurora B is hit by multiple kinase inhibitors, many of which were initially characterized for other unrelated targets (Figure 5). Interestingly, kinases originally defined as classic tyrosine kinase inhibitors have strong effects on serine/threonine kinases and vice versa. Nine kinases in the panel [EphB3, FGFR (fibroblast growth factor receptor) 4, NEK (never in mitosis in Aspergillus nidulans-related kinase) 2, NEK6, Snk, VRK2 (vaccinia-related kinase 2), eEF2K, WNK2 and mTOR] were found to be only weakly inhibited (activity level of 1) by any inhibitors in the collection and only a single kinase, NEK7, was not touched by any of the compounds tested (Supplementary Table S5). In contrast, a larger number of kinases, including popular drug targets such as Flt (Fms-like tyrosine kinase)1–3 and Aurora B, were inhibited by a large number of compounds, and may represent targets that are more sensitive to small molecule inhibition in general (Supplementary Table S5).
Strikingly, we found that there are multiple inhibitors that inhibit previously unidentified targets with high potency. For example, a tyrosine kinase inhibitor, Flt-3 inhibitor II (CAS 896138-40-2) is also active against haspin, a serine/threonine kinase required for normal mitotic progression (Supplementary Table S5, ). The CDK4 inhibitor (CAS 546102-60-7) also hits Flt4 and Aurora B, clinically relevant drug targets from different parts of the kinome tree (Supplementary Table S5). These off-target effects can be seen across the kinome for both serine/threonine and tyrosine kinases. This has significant implications for drug development and polypharmacology strategies which seek to target multiple pathways or multiple targets within a pathway. Furthermore, the interpretation of in-cell or animal data when using small molecules to define signalling mechanism needs to be considered carefully when using small molecules whose activity profiles have not been extensively investigated. We were able to validate many of the originally published inhibitor targets if the target was included in the screen (Supplementary Table S5, Row6 Reported Target Validation, X=validated target, NA=target not screened). However, there are multiple examples where we did not detect inhibition of the reported target. In some cases this may be due to different platforms or systems employed for activity measurements and in other cases it is due to mechanism of the inhibitor. For example, JAK3 (Janus kinase 3) Inhibitor (CAS 211555-04-3) was characterized initially using a gel-based autophosphorylation assay of immunoprecipitates pulled down with a polyclonal antibody for JAK3 . It is difficult to compare results from this type of assay with an assay employing a peptide substrate and purified components including a truncated JAK3 lacking the FERM (4.1/ezrin/radixin/moesin) and SH2 (Src homology 2) domain. It is possible that the JAK3 inhibitor requires full-length JAK3 for inhibition. Another example, PD 98059 (CAS 167869-21-8), a popular MAPK (mitogen-activated protein kinase) activation inhibitor, did not display any inhibitory activity in our screen, presumably due to the fact that its mechanism of inhibition is via binding to the inactive form of MEK [MAPK/ERK (extracellular-signal-regulated kinase) kinase] . Owing to the nature of the recombinant purified kinases used in the screen, the effect of inhibitors on inactive forms of MEK or MAPK could not be evaluated. We discuss a few novel findings of target specificity for important signalling molecules below.
EGFR (epidermal growth factor receptor) inhibitors
Out of the 11 compounds described as EGFR inhibitors, we identified three as the most potent and selective. EGFR Inhibitor (CAS 879127-07-8, S=0.01) and PD 174265 (CAS 216163-53-0, S=0.01) are the most potent and specific compounds in the collection, inhibiting only EGFR with high potency. PD 158780 (CAS 171179-06-9, S=0.02) is also a very potent and selective EGFR inhibitor, although it also inhibited Lyn in our in vitro profile. We only screened EGFR and ErbB-4, and it would be important to determine cross reactivity with ErbB-2 or ErbB-3 heterodimers to see the complete profile across this important drug target family. Nevertheless, it is clear from our profiling that EGFR Inhibitor and PD 174265 are the most useful compounds for interrogating pathways involving EGFR.
VEGFR (vascular endothelial growth factor receptor) inhibitors
Of the 11 VEGFR (Class V) or PDGFR (platelet-derived growth factor receptor) (Class III) tyrosine kinase receptor inhibitors (Table 2), we found that VEGFR2 Kinase Inhibitor I (CAS 1596693-5, S=0.02) and VEGFR2 Kinase Inhibitor II (CAS 288144-20-7, S=0.02) are potent and selective for Flt3 and Flt4, whereas VEGFR Tyrosine Kinase Inhibitor II (CAS 269390-69-4, S=0.06) is a pan-VEGFR inhibitor potently targeting Flt1, Flt4, KDR (kinase insert domain-containing receptor) and FMS. Additionally, GTP-14564 (CAS 34823-86-4, S=0.06) is a pan-VEGFR/Class III tyrosine kinase inhibitor targeting Flt3, Flt4, KDR, KIT and FMS. Additionally, cFMS Receptor Tyrosine Kinase Inhibitor (CAS 870483-87-7, S=0.02) was found to be a very specific FMS inhibitor which did not inhibit other VEGFR or Class III tyrosine kinases. We recommend the use of these compounds when interrogating pathways related to these important signalling targets.
Other examples of inhibitors of an important target include Syk Inhibitor (CAS 622387-85-3, S=0.22), Syk Inhibitor II (CAS 227449-73-2, S=0.01) and Syk Inhibitor III (CAS 1485-00-3, S=0.00). Syk is broadly involved in regulating lymphocyte immune function, mainly through its obligate role in cellular activation in response to receptor engagement of antigens. There are currently good examples of Syk inhibitors in development for treatment of disease related to inflammation . In our profiling screen we found Syk Inhibitor II to be the most potent and selective inhibitor tested for this target (Table 3). On the basis of these in vitro results, we recommend the use of this compound for specific inhibition of Syk.
Out of the seven compounds originally identified as GSK-specific inhibitors, we identify two as the most potent and selective. GSK-3β inhibitor XI (CAS 626604-39-5, S=0.01) is the most potent and specific GSK inhibitor in the collection, inhibiting only GSK3β (potency=2) and GSK3α (potency=3) with high potency and not touching CDKs or other kinases screened, except for Aurora B. Additionally, GSK-3β inhibitor VIII (CAS 487021-52-3, S=0.03) is very potent and selective for GSK3α (potency=3) in our profile, whereas it is less potent against GSK3β (potency=1). In our profile, BIO (GSK-3 inhibitor IX, CAS 667463-62-9, S=0.24) is a very promiscuous compound hitting other targets unrelated to GSK, in line with other published data .
p38 MAPK inhibitors
Compared with the average kinase inhibitor, we find that the p38 MAPK inhibitors SB202190 (CAS 152121-30-7, S=0.04), SB203580 (CAS 152121-47-6, S=0.03) and PD169316 (CAS 152121-53-4, S=0.02) are selective, as reported in the literature (Table 4 and Figure 6, top panels) [26,35,36], potently and specifically inhibiting p38α [SAPK (stress-activated protein kinase) 2a or MAPK14] and p38β (SAPK2b or MAPK11). In Figure 6, potency and selectivity for individual small molecules are visualized in a simple dartboard for each of the 234 kinases. All three p38 MAPK inhibitors display relatively selective profiles (Figure 6, top panel) and S scores in the range of 0.02–0.04. Additionally, all three inhibitors potently inhibit p38α and p38β without touching the other p38 isoforms, γ and δ. In contrast, the results from the present study show that another MAPK inhibitor, SB220025 (CAS 165806-53-1, S=0.03), can differentiate between p38α and p38β, at least in vitro (Supplementary Table S5). We suggest using this compound in parallel with the previously mentioned inhibitors to differentiate the relative contribution of MAPK isoforms within a cellular context. However, as shown in the top panel of Figure 6, we did detect other potent off-target effects that should be considered when using these compounds. Additional follow-up IC50 confirmations and in-cell studies should be performed, since these compounds show potent inhibition of other important signalling kinases. For example, SB203580 potently inhibits EGFR, both SB202190 and PD 169316 potently inhibit CK1δ, and SB202190 additionally hits JNK3, all of which are current drug targets with profound roles in growth signalling, differentiation and cell death responses [37–39].
JNK inhibitors have been used extensively in assigning function to one or all of the JNKs (JNK1, JNK2 and JNK3), especially as they relate to toxic response, apoptosis signalling and transformation [40–42]. From the results of the present study it is clear that there are very selective compounds within our screen suitable for pan-JNK inhibition and JNK3-specific inhibition; however, there is a need for JNK1- and JNK2-specific tool compounds (Table 5 and Figure 6, bottom panels). JNK Inhibitor II, SP600125 (CAS 129-56-6, S=0.18), was one of the first commercially available JNK inhibitors, and has been used in over 2000 published studies which complemented molecular and siRNA (small interfering RNA) studies in showing the importance of JNK1–JNK3 in specific pathways relevant for disease processes related to neurodegeneration, cancer and inflammation [40–42]. However, consistent with previous studies [24,43], it is clear that data generated using the relatively non-specific SP600125 should be carefully evaluated in light of the in vitro data showing potency against 20 other kinases, many of which are also important drug development targets, including Aurora B/C, Flt4, Flt3, TrkA, CK1δ, CDK2 and CDK5. Additionally, we also found that JNK inhibitor V (CAS 345987-15-7, S=0.06) is a potent inhibitor of GSK3α and GSK3β in vitro (Supplementary Table S5) and studies using this compound should be evaluated with respect to this signalling pathway.
ROCK (Rho kinase) inhibitors
Of the four molecules (Y-27632, HA1077, ROCK Inhibitor III and ROCK Inhibitor IV) targeting ROCK-I and ROCK-II, which regulate cytoskeletal signalling and death response due to loss of adhesion , ROCK Inhibitor IV (also called H-1152, CAS 451462-58-1, S=0.04) is the most potent and selective compound (Supplementary Table S5). It is even more potent than Y-27632, the important stem cell reagent which promotes embryonic stem cell and iPS (induced pluripotent stem) cell expansion in vitro . We suggest using ROCK Inhibitor IV when interrogating this target's involvement in signalling.
TGFβRI Inhibitor III (CAS 356559-13-2, S=0.01) and TGFβRI Kinase Inhibitor (CAS 396129-53-6, S=0.02) both potently inhibit TGFβRI (ALK5) and ALK4, with very little effect on other kinases screened. We can recommend both with the former being slightly more selective.
Aurora kinase inhibitors
Of the four Aurora kinase inhibitors in the collection, Aurora Kinase Inhibitor II (CAS 331770-21-9, S=0.04) is the most potent and selective pan-Aurora inhibitor (Aurora A/B/C) with a profile that shows it is most potent for Aurora B (potency=3) followed by C (potency=2) and least potent for A (potency=1). Aurora Kinase Inhibitor III (CAS 879127-16-9, S=0.12) is also potent for Aurora A and B, but does not hit Aurora C. We also discovered an additional selective ‘off-target’ inhibition of Aurora kinases as discussed below.
Other highly specific and potent kinase inhibitors from our screen include, but are not limited to, the following: Chk2 Inhibitor II (CAS 516480-79-8, S=0.01) is highly specific for Chk2 without touching Chk1 and with very little effect on other kinases; CDK/Crk inhibitor (CAS 784211-09-2, S=0.05) is the most selective pan-CDK inhibitor; rapamycin (CAS 53123-88-9, S=0.00), as expected, is highly specific for mTOR.
Kinase target sensitivity
The corollary to inhibitor promiscuity is target sensitivity. That is, just as small molecules can be ranked by their selectivity, kinases can be ranked by their sensitivity to inhibition within the chemical space defined by the subset of compounds being screened. In a manner similar to inhibitor selectivity, we assigned a selectivity score for each of the 234 kinase targets profiled. This measure can be seen as the intrinsic sensitivity of a kinase target to small molecule inhibitors, especially those which are ATP competitive. Different from true druggability, this score may provide information on active site accessibility or flexibility, from both structural and chemical perspectives. Within our profile we found the median number of compounds inhibiting a kinase to be 12 (Figure 4B). In this case, the S’ score is based upon the percentage of small molecules that inhibit a kinase target with high potency (activity level of 2 or higher) among all small molecules screened. Similar to the S score for inhibitors, the kinase targets themselves display a wide range of selectivity, which we have found to be independent of factors such as the Km for ATP (Figure 4B and Supplementary Table S5). For example, Flt4, the most sensitive kinase tested, is inhibited by 69 compounds with activity potency assignments of at least 2 and a predicted IC50 ≪10 μM. We suggest that drug discovery strategies should screen not only close family members and kinases on the basis of protein similarity, but also the kinases which are most sensitive to inhibition. For example the list of the 20 most sensitive targets is over-represented by tyrosine kinases (75%), although the overall collection comprises only 31% tyrosine kinases, and the proportion of inhibitors initially published as targeting tyrosine kinases is 35%. Inhibition of these targets can have pleiotropic cellular effects and should be considered when performing phenotypic screens. For example, pharmacological inhibition of EGFR can lead to downstream JNK activation and apoptosis . This list also includes important cell cycle regulatory targets such as Aurora A and B and CaMKIIδ, an important toxicity target involved in regulation of calcium signalling in the heart . These might represent potent off-target effects that may be deleterious and associated with toxicity or unwanted effects.
On the basis of the above observation, we hypothesized that a smaller subset of kinases may serve as a potential alternative test set for off-target kinome screens to improve efficiency or when a larger screen is prohibitive due to resource constraints. To validate our hypothesis and to determine whether a smaller subset of kinases could be used as a strategy to test for general cross-reactivity, we performed statistical analysis comparing the orders of small molecules ranked by S scores. We compared the rank order when the top 20, 30, 50 and 100 sensitive kinases were used relative to the result with the complete set of 234 kinases by two different measures, simple linear regression correlation coefficient and Spearman's rho rank correlation coefficient. We demonstrate that even a set of 20 kinases are a good representation of the complete set of kinases for ranking the selectivity of small molecules (R=0.71 and rho=0.88, and Supplementary Table S7 at http://www.biochemj.org/bj/451/bj4510313add.htm). Standard testing of these top 20 kinases thus would represent a valid approach to reducing the number of off-target effects during drug candidate selection and optimization. Obviously, screening a larger diverse set of compounds can further optimize the selection of the most sensitive kinase subsets for routine screening purposes. A more stringent test to fully understand the robustness of this approach will involve assaying a completely different set of small molecules that have not been used as the training set for the selection of a sensitive kinase subset.
To cluster the kinase inhibitors on the basis of their activity profiles, we assigned potency across the complete profile. Inhibitors with activity levels greater than or equal to 2 were treated as active, and the others as inactive. Hierarchical clustering of small molecules by activity fingerprints using a binary method allows intuitive exploration of compounds with similar activity profiles (Figure 7A). With additional validation and expansion of the library, such an approach may allow prediction of biological functions based on similar activity profiles. In addition, compounds were also clustered based on their 2D chemical structures (Figure 7B). We used the 2D fingerprints from PubChem for this purpose. Converting chemical structures to 2D fingerprints is a convenient standard method to allow mathematical manipulation of structure information; however, the method cannot fully distinguish certain functional groups and features. Nonetheless, this approach allows rapid informatics analysis of large sets of data with reasonable confidence. Similarity matrixes with all compound pair-wise similarity measurements in terms of kinase activity profile and 2D structures can be found in Supplementary Tables S8 and S9 (at http://www.biochemj.org/bj/451/bj4510313add.htm). These can aid in the identification of potential small molecule leads on the basis of activity profiles rather than chemical structures alone. For example, a similarity matrix can be used to help researchers find the closest compounds in terms of activity (Figure 7C) or structure profile for follow-up screening or explaining experimental observations. The clustering results can be visualized in a similar manner as phylogenetic trees for both activity (Figure 7A, and Supplementary Figure S2 at http://www.biochemj.org/bj/451/bj4510313add.htm) and structure (Figure 7B and Supplementary Figure S3 at http://www.biochemj.org/bj/451/bj4510313add.htm).
Activity profile similarity compared with structural similarity
From our profile analysis, compounds with very similar activity profiles (similarity >0.8) tend to have very similar structures. For example, compound pairs that are similar in terms of promiscuity, such as staurosporine and its close analogue K-252a, and compound pairs that are more selective, such as SB 203580 (CAS 152121-47-6) and its close analogues (Figure 8A, top panel) share structural similarity. Interestingly, although they share high activity profile similarity (similarity >0.8), SB 203580 and p38 MAPK Inhibitor (CAS 219138-24-6) are of different chemical scaffolds (Figure 8A, middle panel). Likewise, PD 169316 (CAS 152121-53-4) and SKF-86002 (CAS 72873-74-6) share activity profile similarity, but different chemical scaffolds (Figure 8A, bottom panel). These compound pairs inhibit shared kinase targets with an activity level greater than or equal to 2. This is especially interesting from a medicinal chemistry perspective, as it implies that it is possible to design compounds with desired polypharmacology activity profiles from distinct chemical scaffolds.
Comparing activity profiles to structure similarity also highlights the fact that similar structures do not always result in similar activity profiles. In fact, many structurally similar compound pairs have divergent activity profile similarities. For example, ERK Inhibitor II, FR180204 (CAS 865362-74-9) and ERK Inhibitor II, Negative Control (CAS 1177970-73-8) in Figure 8(B) have very similar structures, but very different activity profiles and do not share any targets with activity levels greater than or equal to 2. The former is active against JNK2α2, JNK3, MAPK1 and PKCδ (protein kinase Cδ), whereas the later is only active against Aurora B and IRR (insulin receptor-related) (Figure 8B and Supplementary Table S5).
As expected, we find many examples of compound pairs with similar activity profiles and similar structures, as well as compounds with similar structures, yet different activity profiles. However, the most unexpected results come from two pairs of compounds with similar activity profiles (similarity >0.65), yet very different structures. The first compound pair is PI3Kγ (phosphoinositide 3-kinase γ) Inhibitor II (CAS 648449-76-7, active for CK2, CK2a2, Pim-1 and Pim-2) and CK2 Inhibitor III, TBCA [CAS 934358-00-6, active for CK2, CK2a2, PASK, Pim-1, Pim-2 and Plk3 (Polo-like kinase 3)] (Figure 8C, top). Another example is VEGFR2 Kinase Inhibitor II (CAS 288144-20-7, active for Flt1, Flt3, Flt4, Fms and KDR) and AGL 2043 (CAS 22617-28-8, active for Flt3, Flt4, Fms, KDR and Aurora-B) (Figure 8C, bottom panel). These are intriguing comparisons, which may represent interesting bioisosteres and/or slightly different binding modes [48,49]. For example, the thiazolidinedione moiety found in PI3Kγ inhibitor II can be considered an isostere of the carboxylic acid functional group that is present in CK2 Inhibitor III, TBCA.
The use of small molecules as tools for exploring protein function and its relationship with signalling, cellular phenotype and disease has become a ubiquitous strategy within the biological community. Beginning with the well-developed kinase target area, we have performed studies to provide additional selectivity and potency information related to commonly used research reagent compounds. Such information demonstrates how small-molecule inhibitors should best be used in defining casual and functional relationships for pathway analysis and drug development strategies related to polypharmacology. In some cases, a larger set of chemicals will be needed to more extensively interrogate the kinome; nevertheless, our set of data provides a framework to build upon and is useful in identifying potential off-target effects, especially for those targets that might be associated with toxicity or other unwanted properties.
Potency and selectivity profiles can be used to identify appropriate compounds or sets of compounds for use in determining functional necessity for a gene or protein within a particular pathway, cellular function or pathology. We have already discussed above the use of some popular inhibitors of targets important for disease and signalling. Using the JNK inhibitor SP600125 as an example, we have identified that SP600125 inhibits haspin and DYRK2 (dual-specificity tyrosine-phosphorylated and -regulated kinase 2), the analysis of which may provide novel insights into possible mechanisms of action of SP600125. Similar to Aurora B and C, which are inhibited with high potency, haspin is emerging as a mitotic signal co-ordinator responsible for phosphorylation of histone H3 at Ser10 and provides a docking site for the recruitment of Aurora B to centromeres in a survivin-dependent manner. It is easy to imagine that many of the G2 effects associated with SP600125  may be related to this activity and not to direct inhibition of any of the JNKs, as inhibition of haspin results in mitotic catastrophe and apoptosis under certain conditions . Additionally, previous informatics analysis suggests that the main effects of SP600125, from a gene expression perspective, are related to changes in the Myc transcription network . This could be an indirect result of JNK inhibition, but is probably more likely due to direct effects on the GSK priming function of DYRK2. Specifically, both c-Myc and Jun are phosphorylated by DYRK2, thereby generating a preferred substrate for GSK3β, a required step for ubiquitin-dependent degradation of these substrates to allow passage through G1 . Inhibition of DYRK2 either by siRNA or small-molecule inhibitors results in tumours, shortening of G1, and driving cells into cycle. Furthermore, DYRK2 expression is reduced in multiple human tumour samples, and down-regulation of DYRK2 correlates with human breast cancer invasiveness and gastric cancer . As the substrates for DYRK2 become better defined, it will be important to determine their influence on cellular processes, especially those related to growth and apoptosis. Thus one of our findings is that DYRK2 is an important kinase to evaluate in profiling screens where cell-cycle-related effects are expected.
Additionally, using the data, one can mine new uses for existing molecules similar to drug development strategies for repurposing existing clinical inhibitors. For example, we were able to detect a number of instances of novel targets that are different from the originally identified target cited in the literature (Table 6). For example, IRAK (interleukin-1-receptor-associated kinase)-1/4 inhibitor (CAS 509093-47-4), a very selective compound, is most potent against the transmembrane tyrosine kinase receptor Fms, important for monocyte and macrophage proliferation and differentiation , with very little effect on IRAK1 or IRAK4 in our screen. Discrepancies in results from the original screen that identified this compound may be explained by the different screening platforms employed. Our radioactive activity assays used MBP (myelin basic protein) as a substrate for both IRAK1 and IRAK4, whereas the original screen used a peptide from GFAP (glial fibrillary acidic protein) as a substrate in a chemilumenescent antibody-based ELISA type assay . In another example, we found SKF-86002 (CAS 72873-74-6), a very specific p38 MAPK inhibitor , to also potently inhibit CK1δ. Another case involves MNK1 (MAPK-interacting kinase) Inhibitor (CAS 522629-08-9), which has been characterized by its ability to inhibit phosphorylation of eIF4E (eukaryotic initiation factor 4E) and abrogate translation initiation in response to proliferative signals . MNK1 inhibitor also potently and selectively hits the cytoplasmic tyrosine kinase Blk, which is required for B-cell signalling as well as the response to insulin [58,59].
Even molecules originally identified as inactive for a specific target and used as a negative control may have other off-target effects that have not yet been explored. For example, the negative control for ERK Inhibitor II (catalogue number 328008), with a substituted OH group attached to the 1H-pyrazolo[3,4-c]pyridazin core rather than an amino group, is inactive against ERK1 and ERK2. Indeed, this molecule does not touch ERK1 or ERK2, but potently and specifically inhibits Aurora B with no effect on Aurora A or Aurora C, and represents the most selective isoform-specific Aurora B tool compound that we are aware of. In fact, this molecule turns out to be the most potent and specific Aurora B inhibitor in the collection. Certainly, our findings call for more in-depth follow-up both in vitro and in cells to determine whether these initial observations can be used for repurposing tool compounds.
An excellent example of how such data can be used to explore biology is related to the single kinase target identified for Akt inhibitor X (CAS 925681-41-0). Of the 234 kinases in our screen, only PASK was inhibited by this molecule. PASK is an emerging target shown to be necessary for regulation of metabolic stress across phylogeny and has been recently proposed to be a potential therapeutic target for diabetes and cancer, in the same class as AMPK and mTOR . It was recently reported that overexpression of either of the two yeast PASKs, PSK1 or PSK2, suppressed a TORC2 (target of rapamycin complex) temperature-sensitive mutation, suggesting that PASKs lie in the same or parallel pathway as mTOR . Interestingly, PASK substrates include translational regulators and overlap with important mTOR substrates . Additionally, Akt Inhibitor X, also known as NCP10, was more recently identified in a cell-based screen for autophagy inducers in primary neuronal cultures and was shown to be a potent inducer of autophagy, similar to mTOR-dependent rapamycin induction of autophagy in other cell types . In order to verify PASK as a hit for Akt Inhibitor X, we followed up with a nine-point determination of the IC50 of this compound for inhibiting PASK. We calculated an IC50 of 3.27 μM (results not shown) for PASK inhibition, similar to the IC50 for both autophagy induction and inhibition of AKT phosphorylation in cells. It is intriguing to imagine that PASK is an important node for autophagy regulation and that Akt Inhibitor X induces autophagy via PASK activity inhibition. The in vivo relevance of our finding remains to be determined with further experiments in cells.
Finally, the comparison of activity profiles with chemical structures demonstrates a novel strategy for polypharmacology-driven drug discovery strategies. For example, our results imply that it may be possible to design compounds with desired polypharmacology using distinct chemistry scaffolds. If these compounds possess similar drug metabolism and pharmacokinetics properties, they may presumably have similar in vivo pharmacological profiles. This is especially encouraging from a medicinal chemistry perspective, since it is often necessary to fine tune the chemistry scaffold to optimize ADME (absorption, distribution, metabolism and excretion) properties or to create novel intellectual property during drug discovery. It is well documented that scaffold ‘hopping’ is possible when designing compounds that are active against certain specific biological targets, but it is not as obvious that compounds derivatized from different templates can have very similar activity profiles against a very large set of targets. This may represent a new approach to drug design for polypharmacology.
In summary, we have shown the value in performing larger profiling screens encompassing kinase targets from across the kinome. The fact that less than 10% of the kinome is being targeted in current clinical trials for oncology and that the majority of all clinical kinase inhibitors are for targets with an approved drug  suggests that new drug target selection could be addressed by a systematic evaluation of the kinome. Similar to other recent large kinase profiling studies, we provide a framework for assessing polypharmacology using the in vitro inhibition profiles of 158 inhibitors against 234 human kinases. For example, it is clear from our selectivity and potency profiles of JNK inhibitors that results using first-generation inhibitors such as SP600125 are difficult to interpret in the light of our findings that numerous cell cycle regulators and growth factor receptors are also inhibited. Furthermore, it is clear that some kinases are more susceptible to small-molecule inhibition and that routine off-target profiling of these kinases may be warranted. Finally, we describe potential ‘off-target’ selective inhibition of Blk, Aurora B, CK1δ and Fms for current tool inhibitors. We believe that these observations represent only a fraction of the useful knowledge generated from such profiling data, and we presume that the research community will be able to mine more from the data that we provide. Such data can be an important resource for both drug development and basic research and should aid in the identification of potential off-target issues, identification of multi-kinase inhibitors, and quick access to lead candidates for new targets.
Kevin Harvey and Yinghong Gao conceived the experiments. Stephen Davies and Anna Woodward executed the biochemical assays. Yinghong Gao executed all data analysis. Kevin Harvey wrote the paper. Umesh Patel, Martin Augusitin and Robert Koveleman provided oversight, editorial and scientific review of the project and paper prior to submission.
All authors are employees of EMD Millipore, a division of Merck KGaA, Darmstadt, Germany.
We thank Dominique Perrin, Eike Staub, Michael Krug, Blaine Armbruster, Fabio Ferfoglia, Lina Yip Sonderegger and Alexander Scheer for useful discussion.
Abbreviations: 2D, two-dimensional; AMPK, AMP-activated protein kinase; CaMK, Ca2+/calmodulin-dependent protein kinase; CAS, Chemical Abstract Service; CDK, cyclin-dependent kinase; DYRK2, dual-specificity tyrosine-phosphorylated and -regulated kinase 2; eEF2K, eukaryotic elongation factor 2 kinase; EGFR, epidermal growth factor receptor; ERK, extracellular-signal-regulated kinase; FGFR, fibroblast growth factor receptor; Flt, Fms-like tyrosine kinase; GSK, glycogen synthase kinase; IRAK, interleukin-1-receptor-associated kinase; JAK3, Janus kinase 3; JNK, c-Jun N-terminal kinase; KDR, kinase insert domain-containing receptor; MAPK, mitogen-activated protein kinase; MEK, MAPK/ERK kinase; MNK1, MAPK-interacting kinase 1; mTOR, mammalian target of rapamycin; NEK, never in mitosis in Aspergillus nidulans-related kinase; PASK, Per/Arnt/Sim domain-containing serine/threonine-protein kinase; PI3Kγ, phosphoinositide 3-kinase γ; ROCK, Rho kinase; SAPK, stress-activated protein kinase; siRNA, small interfering RNA; TGFβR, transforming growth factor-β receptor; VEGFR, vascular endothelial growth factor receptor
- © The Authors Journal compilation © 2013 Biochemical Society