Promiscuity analysis of a kinase panel screen with designated p38 alpha inhibitors

Abstract
Protein phosphorylation by kinases is of critical importance for the regulation of many cellular functions. When kinases are deregulated numerous biological processes are affected, which may cause a variety of diseases. Therefore, kinase inhibition plays an important role in therapeutic intervention. A number of kinase inhibitors have been approved as drugs, initially in oncology where promiscuous (multi-kinase) inhibitors were most efficacious. Exploring kinase inhibitor selectivity and promiscuity for therapy is among the most challenging aspects of kinase drug discovery. Herein, we thoroughly analyze a kinase profiling experiment in which 637 designated inhibitors of p38 MAP kinase (p38α) were tested against a panel of 60 kinases distributed across the human kinome. In this experiment, only 19% of the inhibitors were found to be promiscuous when the median p38α inhibition rate was applied as an activity threshold. Promiscuous inhibitors had a median value of two targets per compound, and many of these inhibitors were only active against the p38α and closely related JNK3 enzymes. Promiscuity cliffs were identified and analyzed in a network representation revealing structural modifications that were implicated in triggering compound promiscuity. Taken together, the findings revealed a high degree of selectivity of designated p38α directed inhibitors although they target the ATP binding site that is largely conserved across the human kinome.

1.Introduction
Kinases play a critically important role in modulating cellular processes such as differentiation, proliferation, apoptosis, metabolism, and regulation of immunological responses [1, 2]. Abnormalities in kinase signaling pathways are implicated in a variety of disease states including cancer [3] and autoimmune diseases [4]. Kinase inhibitors have become promising therapeutic agents for the treatment of such diseases [5] and a number of kinase inhibitors have been approved as drugs, especially in oncology [5-7]. Most kinase inhibitors developed to date target the adenosine triphosphate (ATP) cofactor binding site, which is largely conserved across the kinome, or they target regions proximal to the ATP site [6-8]. Therefore, such inhibitors are often expected to have multi-kinase activity, which is also referred to as promiscuity [9]. Inhibitor selectivity on the one hand and promiscuity on the other play an important role for therapeutic applications, for which kinase inhibitors provide excellent examples. In cancer treatment, promiscuous kinase inhibitors that are capable of simultaneously interfering with multiple signaling pathways have been proven to be most efficacious [6]. By contrast, for other therapeutic applications such as the treatment of chronic inflammatory diseases, inhibitors typically require a high degree of kinase selectivity [8]. Kinase inhibitor selectivity versus promiscuity and implications for therapy continue to be intensely debated. These issues are often controversially viewed and are currently not fully understood [10-13].

For example, although ATP binding site directed kinase inhibitors are generally expected to be promiscuous, given the conservation of the ATP site, such inhibitors often display different degrees of promiscuity and also include kinase-selective compounds [13, 14]. Binding characteristics of kinase inhibitors have mostly been evaluated using profiling experiments in which inhibitors are screened against panels of kinases, yielding matrices of inhibitor-kinase interactions [14-17]. Such panel screens have become a major source of selectivity vs. promiscuity data for inhibitors across the kinome.Profiling experiments had a major impact on the kinase field because they made it possible to examine binding characteristics of kinase inhibitor under given experimental conditions in context and provided comprehensive views of promiscuity versus selectivity of inhibitors. For ATP site directed inhibitors, unexpected selectivity trends were detected early on[14] and in subsequent investigations [15,16], which made important contributions to rationalizing the interplay between selectivity and promiscuity as wide spectrum of related yet distinct binding events. Experimental profiling was complemented by computational analysis of kinase inhibitor activity data from biological screening and medicinal chemistry [9,13]. Systematic calculations detailed compound coverage of the kinome and provided global data- driven views of selectivity and promiscuity that helped to balance general expectations [9].

Our study has focused on designated inhibitors of p38 MAP kinase (p38α), which is strongly implicated in angiogenesis, apoptosis, and cancer [18-20]. The p38α enzyme belongs to the MAP kinase family comprising four members with differences in expression patterns, substrate specificity, and sensitivity to small molecules. The MAP kinase pathway is triggered by cellular stress conditions as well as cytokines and is responsible for normal and inflammatory immune responses. It primarily regulates the production of pro-inflammatory cytokines, but is also implicated in many other regulatory mechanisms such as cell cycle control. In the MAP kinase family, p38α has become a prominent target for inhibiting inflammatory responses. In addition, p38α has also become a well-recognized target for anti-cancer therapy because p38α signaling is also implicated in supporting cancer cell proliferation [20].We report a detailed promiscuity analysis of a kinase profiling experiment in which 637 designated ATP site directed inhibitors of p38α were screened against a panel of 60 kinases distributed across the human kinome. Our analysis focused on the question how selective or promiscuous kinase inhibitors might be that were originally designed to target p38α.

2.Experimental
The 637 unique compounds resulted from in-house synthetic efforts representing different generation of p38α inhibitors. The compounds belonged to different structural classes including, among others, pyridinylimidazoles, dibenzosuberones, pyridinyloxazoles, purines, doxepinones, pyranopyrazoles, triazolo-thiadiazoles, as illustrated in Figure 1a. The composition of the compound collection is summarized in Table 1. All compounds were synthesized, characterized, and purified, as described previously [21-32]. With 271 instances, pyridinylimidazole derivatives were the prevalent class of inhibitors. Figure 1b shows the schematic representation of the X-ray structure of a complex between p38α and a representative pyridinylimidazole inhibitor (SB203580) [33], illustrating interaction patterns within the ATP binding site.Several structural features of p38α support selective inhibition over other kinases [25,33]. The presence of a small gatekeeper residue (Thr106) controlling access to a hydrophobic binding area makes it possible for compounds with linearly arranged aromatic rings to closely interact with the hinge region of p38α. In addition, canonical residue Met109 is followed by Gly110. This pairing is only present in less than 10% of the kinome. The small Gly110 residue is flexible and can rotate, forming the so-called glycine flip, which enables the formation of dual hydrogen bonds with carbonyl oxygen containing compounds through two proximal amino group donor functions. Accordingly, combination of the small gatekeeper residue and the glycine flip is regarded as a major inhibitor selectivity determinant of p38α.The 637 designated p38α inhibitors were screened against 60 different human kinases reported in Table 2. Kinase profiling was carried out using the Caliper mobility shift assay format as reported previously [34].

In preliminary experiments, exemplary p38 inhibitors were tested at different single concentrations in p38 and other kinase assays to reproducibly distinguish active from inactive measurements. On the basis of these test calculations, a single compound concentration of 5.5 M was used for profiling and the percentage of inhibition was determined as an activity measurement. Promiscuity is defined herein as the ability of compounds to specifically interact with multiple kinase targets [9]. As a measure of promiscuity, the promiscuity degree (PD) [35] of a compound was calculated as the number of kinases the inhibitor was active against on the basis of a specific percentage of inhibition threshold. Thus, the PD accounts for all activities of inhibitors across the kinome.To assess promiscuity relationships between structural analogs, the matched molecular pair (MMP) formalism was applied [36]. An MMP is defined as a pair of compounds that only differ by a chemical change at a single site, which corresponds to a single-site substitution yielding two structural analogs. The substitution can be rationalized as an exchange of a pair of substructures, which is termed a chemical transformation [36]. For the 637 test compounds, transformation-size restricted MMPs [37] were systematically calculated in whichtransformations were limited to chemical modifications typically observed in series of analogs [37].

The 637 compounds yielded a total of 2526 unique MMPs that involved 548 compounds. MMP calculations were carried out with an in-house implementation of the Hussain and Rea algorithm [36].MMPs capturing large PD differences (as further specified below) were classified as “promiscuity cliffs” (PCs) [38]. A PC is defined as a pair of structurally analogous compounds with a large difference in their PD value. It requires the presence of a non- or weakly promiscuous compound and a highly promiscuous analog. Thus, PCs capture small structural modifications that are implicated in compound promiscuity [38]. PCs were organized and visualized in PC network representation in which nodes represent compounds from PCs and edges pairwise PC relationships the compounds form [39]. In PC networks, clusters typically emerge as disjoint network components. The clusters are mostly formed by overlapping PCs that share PC compounds. In these clusters, PC pathways (PCPs) [39] can be algorithmically traced[40] that consist of sequences of PCs with alternating highly promiscuous and weakly or non- promiscuous compounds and are thus rich in SAR information. PC, PC network, and PCP calculations were carried out with in-house implemented software tools and algorithms (for PCP identification and tracing).

3.Results and discussion
By design, compounds evaluated in the panel screen preferentially inhibited p38α. Therefore, the distribution of percentages of inhibition of p38α was used to determine thresholds for binary classification of compounds as active or inactive. Figure 2 reports the distribution of the percentage of p38α inhibition for all test compounds. Most compounds inhibited p38α at the level of 40% and above. The median value of the distribution was 86.8% inhibition. Hence, about half of the test compounds inhibited p38α at this level or above. Therefore, a percentage of inhibition ≥ 86.8% was set as a general threshold for inhibitory activity of compounds. As control thresholds for our analysis, the mean of the distribution (68.6% inhibition) as well as 50% of inhibition were also used.With each of the three percentage of inhibition thresholds, inhibitors of the 60 kinases were identified. At the most stringent threshold (percentage of inhibition ≥ 86.8%), 335 compounds (52.5%) were found to inhibit at least one kinase. Hence, at this level, about half of the compounds were classified as inactive. Figure 3a shows the resulting distribution of inhibitors across the human kinome. Following p38α, the largest number of inhibitors was found for the closely related JNK3 kinase (94 inhibitors) and MAPK4 (42 inhibitors).

JNK3 and p38α are two homologous protein-serine/threonine kinases belonging to the same CMGC kinase groupand to the mitogen-activated protein kinase (MAPK) family [41]. MAP4K4 is a member of theSTE group but also participates in the MAPK signaling pathway and is directly involved in JNK activation [42]. However, rather different results were obtained for other members of the CMGC group. For example, none of the assayed compounds inhibited CDK1, CDK2, or CDK4 and only one inhibitor of Erk2 was identified. By contrast, kinases with 10 to 25 inhibitors included EGFR, KDR and ErbB4 (TK group) as well as TNIK from the STE group. Only a single or no inhibitor was found for the remaining kinase groups.For the percentage of inhibition ≥ 68.8% threshold (i.e., the average p38α inhibition level), 429 compounds (67.3%) were classified as inhibitors. The corresponding kinome map is shown in Figure 3b. Thus, compared to the median inhibition rate, an increase in the proportion of active compounds of ~15% was observed. For the percentage of inhibition ≥ 50% threshold, 475 (74.6%) of the compounds were classified as active (Figure 3c), yielding only another slight increase of ~7%. Hence, the number of weakly active compounds was overall limited and there was no substantial increase in kinome coverage. Rather, the kinome inhibition profile was similar in all three instances, as revealed in Figure 3. Accordingly, the activity data, albeit resulting from single concentrations, were expected to provide a meaningful basis for promiscuity analysis.

Inhibitors with a PD  2 (i.e., activity against at least two kinases) were classified as promiscuous. For the increasing inhibition thresholds, 285 (44.7%), 219 (34.4%), and 122 (19.2%) promiscuous inhibitors were detected. Thus, for the percentage of inhibition ≥ 86.8% threshold, only every fifth inhibitor was active against at least two kinases. Figure 4 reports the distribution of promiscuous inhibitors for all three thresholds. The distributions were overallnarrow and promiscuity degrees were generally low. Only a few “statistical outliers” (dots above the upper whiskers) were observed, representing the most promiscuous compounds that were detected. For the increasing inhibition thresholds, median PD values of 4, 3, and 2 (with a mean of 3) were obtained, respectively. At the percentage of inhibition ≥ 86.8% threshold, 52 of the 122 promiscuous inhibitors were exclusively active against the closely related p38α and JNK3. Thus, only 70 inhibitors displayed other off-target activities against more distantly related kinases.Figure 5 reports the distribution of promiscuous inhibitors over different PD values for the percentage of inhibition  86.8% threshold, which further illustrates the presence of overall low promiscuity. Figure 6 reports inactive, selective, and increasingly promiscuous inhibitors from different structural classes for the percentage of inhibition  86.8% threshold.

The highest observed PD value was 17 for a single inhibitor from the pyridinylpyrazole structural class. More than half of the tested inhibitors from the dibenzosuberone class were active and selective, whereas inhibitors from the triazolo-thiadiazole class were mostly inactive. Furthermore, pyridinylpyrazole inhibitors were either inactive or promiscuous with PD ≥ 5. Only two inhibitors had a PD > 10, one of which was a pyridinylimidazole and did not inhibit p38α. Thus, different structural classes displayed different kinase inhibition characteristics and varying proportions of inactive, selective, and promiscuous compounds. Overall, at the inhibition  86.8% threshold, 302 (47.5%) compounds inactive against all 60 assayed kinases and 207 (32.5%) inhibitors were only active against p38α. Thus, under the experimental conditions of the panel screen, about a third of the compounds were selective for p38α, their primary target.We identified inhibitors belonging to different structural classes that were only active against a single kinase at all three inhibition thresholds. Applying this stringent selectivity measure, a total of 89 selective inhibitors were identified for the majority of structural classes. None of the purine and pyridinylpyrazole inhibitors were found to be selective. As reported in Table 3, varying numbers of selective inhibitors were identified for different classes. Notably, all of these inhibitors were selective for p38 but no other kinase. Most of the selective inhibitors were pyridinylimidazoles, dibenzosuberones and doxepinones, with 22, 17, and 13 compounds, respectively.

In addition, 27 inhibitors selective for p38 originated from other structural classes with less than 10 compounds per class. Figure 7 shows exemplary inhibitors.The PC data structure aids in the identification of small chemical modifications accompanying significant changes in multi-target activities. On the basis of the PD value distribution for the percentage of inhibition ≥ 86.8% threshold, compounds considered to be “highly promiscuous” in our profiling experiment were required to have a PD equal to or greater than the mean plus one sigma (standard deviation), giving rise to PD ≥ 5. In addition, considering that many MMPs (analog pairs) were formed between compounds with PD = 5 and PD = 0 (inactive), the required ∆PD for PC formation was set to ≥ 5. Furthermore, weakly promiscuous compounds were confined to those having a PD value equal to or smaller than the mean of the distribution, i.e., PD ≤ 3. On the basis of these criteria, 110 PCs were obtained that involved 83 unique inhibitors, i.e., only 13.0%. The small proportion of tested compoundsinvolved in the formation of PCs was another consequence of overall low promiscuity. PCs were further analyzed in a network representation shown in Figure 8a. Network analysis revealed that the formation of PCs was centered on four highly promiscuous compounds in the major network component (largest cluster) that were active against six to nine kinases and formed a variety of PCs with weakly or non-promiscuous structural analogs. In PC networks, such compounds representing densely connected nodes have been termed “promiscuity hubs” [39]. The largest cluster in the kinase inhibitor PC network containing such hubs was formed by pyridinylimidazole analogs.

From this cluster, a PCP was isolated that contained all four hubs (Figure 8b) and revealed a number of structural modifications leading to large differences in promiscuity. These observations provide opportunities for follow-up investigations to further explore apparent structural kinase promiscuity determinants. Notably, one of the three small clusters (cluster I at the bottom of Figure 8a) was formed by pyrimidinylindole analogs and displayed a distinct composition. Here, PCs were formed by a consistently inactive compound and four structural analogs that each inhibited five or more kinases (Figure 8c). In this case, the promiscuous inhibitors were distinguished from the inactive analog by the addition of different halogen atoms at corresponding ring positions. Provided the inactive analog was not structurally compromised under experimental conditions (which might also explain consistent inactivity), these well-defined chemical modifications uncovered another interesting structure-promiscuity relationship. Hence, the analysis of PCs and their relationships in PCPs showed that experimentally detected differences in kinase promiscuity could be associated with substitution patterns distinguishing promiscuous from non-promiscuous inhibitors. Corresponding PC networks were generated for the ≥ 68.6% and ≥ 50% inhibition thresholds and are shown inFigure 9 and in Figure 10, respectively, together with detailed views of individual network components.

4.Conclusions
Herein, we have reported promiscuity analysis of kinase inhibitors on the basis of a profiling experiment in which 637 designated p38α inhibitors was screened against a panel of 60 kinases belonging to different groups. Since the inhibitors were directed against the ATP binding site in p38α that is largely conserved in many kinases, it was of particular interest to investigate if -and to what extent- they might exhibit multi-kinase activity. Initially, experimental data were collected and analyzed on the basis of different inhibition thresholds. The analysis was taking into account that the percentage of inhibition measurements were based on a single compound concentration. Hence, resulting activity data were approximate in nature, less accurate than dose- response experiments, and only comparable on a relative scale. However, at decreasing thresholds, kinome maps revealed similar inhibition profiles and only small increases in weakly active compounds were detected. These observations indicated that the data were of sufficient quality for comparative promiscuity analysis. To these ends, a preferred threshold (percentages of inhibition  86.8%) was applied, representing the median value of p38α inhibition. At his level, ~32% of the tested compounds exclusively inhibited p38α.

Moreover, only 122 inhibitors, i.e., less than 20%, were promiscuous, with a low median PD of 2. Nearly half of these compounds only inhibited p38α and JNK3. By contrast, inhibitors with activity against five or more kinases were only rare. Further analysis of inhibitors with varying PDs using the PC, PC network, and PC pathway data structures revealed some well-defined structural modifications that distinguished promiscuous from non-promiscuous compounds detected in the panel screen. A key finding of our analysis has been that promiscuity of the ATP site directed kinase inhibitors investigated herein was overall only low, even on the basis of approximate experimental data. Indeed, a large number of inhibitors were specific for their intended primary target under the conditions of the panel screen, lending credence to VX-745 underlying design approaches and showing that promiscuity of ATP site directed kinase inhibitors cannot be generally assumed. Accordingly, there is considerable potential to generate ATP site directed inhibitors with kinase selectivity for therapeutic applications.