Allotargeting of the kinase domain: insights from in silico studies and comparison with experiments

Abstract

The protein kinase domain is a broadly explored target for drug discovery, despite limitations imposed by its high sequence conservation as a shared modular domain and the development of resistance to drugs. One way of addressing those limitations has been to target its allosteric sites, shortly called allotargeting, in conjunction with, or separately from, its conserved catalytic/orthosteric site that has been widely exploited. Allosteric regulation has gained importance as an alternative to overcome the drawbacks associated with the indiscriminate effect of targeting the active site, and it turned out to be particularly useful for these highly promiscuous and broadly shared kinase domains. Yet, allotargeting often faces challenges as the allosteric sites are not as clearly defined as its orthosteric sites, and the effect on the protein function may not be unambiguously assessed. A robust understanding of the consequence of site-specific allotargeting on the conformational dynamics of the target protein is essential to design effective allotargeting strategies. Recent years have seen important advances in in silico identification of druggable sites and distinguishing among them those sites expected to allosterically mediate conformational switches essential to signal transmission. The present opinion underscores the utility of such computational approaches applied to the kinase domain, with the help of comparison between computational predictions and experimental observations.

Introduction

Protein kinase domains are one of the most widely studied drug targets. Because phosphorylation is an essential posttranslational modification that controls the communication of signals essential to cellular functions or interventions, more than 70 small molecule kinase inhibitors have received FDA-approval, including several in the last couple of years [1–7].

Most small molecule kinase inhibitors bind to the catalytic (orthosteric) site of the kinase domain where they compete with the substrate ATP [8–10]. However, the high sequence and structural conservation of this site complicates the development of specific inhibitors; and many inhibitors promiscuously affect a variety of structural homologs as well as multiple cellular pathways. An alternative, that has gained importance in recent years, is allotargeting, i.e. targeting an allosteric site involved in specific interaction, thus enabling the selective modulators of specific cellular pathways [11,12]. Allosteric drugs have also emerged as viable solutions to alleviate, if not counter, the development of drug resistance associated with targeting the catalytic site of kinase domains. For example, the myristic acid (MYR) binding pocket in Bcr-Abl kinase is an allosteric site whose targeting by a small molecule (GNF-2) together with the targeting of the catalytic pocket by ATP-competitive drugs emerged as a viable strategy for combatting the recalcitrant drug-resistant mutation T315I [13–17]. Likewise, the combination of a GNF-2 analog (ABL001, asciminib), that binds the MYR pocket in Abl and the ATP-competitive inhibitor nilotinib led to complete eradication of Chronic Myelogenous Leukemia (CML) xenografts [18]. Most recently, Lou et al, linked an orthosteric Abl inhibitor (dasatinib) with asciminib to form an extremely potent and specific bitopic inhibitor [19].

Recent years have seen significant progress in in silico characterization of allosteric regulation, as well as identification of key sites involved in allostery [20–25]. In parallel, druggability simulations have proven to accelerate the identification of druggable sites and the construction of pharmacophore models (PMs) for those sites, which in turn, helped retrieve either novel hits or repurposable drugs from libraries of small molecules or drug databases [26–31]. Also, a machine learning approach has been introduced to design kinase inhibitors [32].

Notably, not all druggable sites necessarily affect/alter the function. New methods have been developed for determining the so-called essential sites that cooperatively alter the target proteins’ dynamics and thereby function if targeted by small molecule modulators [33]. The current opinion focuses on the application of these computational methods to kinase domains toward identification of druggable and allosteric sites. While many computational studies have helped elucidate the conformational changes of kinase domains in response to ligand binding [25,34–37], the current opinion highlights the utility of recent in silico methods for predicting critical binding sites (orthosteric and allosteric), and further shows how these methods can help explain the molecular basis of the impact made by these inhibitors on allosteric dynamics and therefore the function of kinase domains.

The kinase domain is a widely shared and structurally conserved modular unit, with well-studied activation/deactivation mechanisms

Two structural hallmarks define the active conformation of the kinase domain (Fig. 1a): a salt-bridge between K295 and E310 (chicken c-Src numbering) stabilizes the N-lobe, and D404 (of the DFG motif) faces into the active site where it coordinates Mg²⁺/ATP [38]. Deactivation takes place either by disrupting the salt-bridge E310 forms with K295 (teal diagram in Fig. 1b) [39]; or by the reorientation of D404 away from the active site (violet diagram in Fig. 1c) [40].

Figure 1. Conserved structure of the kinase domain: Comparison of the active and inactive states illustrated for Src kinase domain.

(a) Ribbon diagram of Src kinase in the active conformation (PDB: 3DQW) [38]. Important residues are shown in sticks and the N- and C-lobes are shown in semi-transparent surface representation in cyan and yellow, respectively. ATPγS at the ATP binding site is shown in white spheres. (b) Src in the inactive, αC-helix-out / DFG Asp-in conformation (PDB: 2SRC) [39]. The red arrow indicates the change in the orientation of E310 side chain compared to that in the active form. (c) Src kinase in the inactive αC-helix-in / DFG-Asp-out conformation (PDB: 2OIQ). The change in D404 orientation with respect to the active form is indicated by the red arrow.

Druggability and allotargeting of the kinase domain: computations point to experimentally observed and additional sites

Kinases represent a widely studied family of target proteins whose function can be modulated by both orthosteric and allosteric drugs, and/or their combinations [11,41]. Fig. 2a illustrates various drugs, orthosteric (ATP-competitive) [38–40,42–46] and allosteric, bound to the kinase domain. Allosteric drugs include those binding to the MYR pocket for Abl kinase (PDB: 3K5V) [17,46], the PIF site for PDK1 (PDB: 4AW1) [44], the peptide-substrate binding site for RTKC8 (PDB id: 8E4T) [47], and the G-loop site for Src kinase [42].

Figure 2. Comparison of drug-binding sites on kinase domain observed in experiments and those predicted by druggability simulations and essential site analysis.

The ligand binding sites of the kinase domain are shown, as detected by experiments (a), and as predicted by computations (b-c). The top and bottom diagrams in each panel display the structure from two perspectives differing by a 90° rotation as indicated. In panel a, Src (PDB: 3DQW) [38] is shown in orange, ATP-bound PKA in yellow (PDB id: 1ATP) [38,46], ABL with ligand at myristate (MYR) site in green (PDB: 3K5V) [17,45], ATP-bound PDK1 with ligand at the PIF site in blue (PDB: 4AW1) [44], and RTKC8 with ligand at peptide-substrate site in salmon (PDB id: 8E4T) [47]. All five structures have ligands bound to the ATP site. Src with ligand at the G-loop site is shown in grey [42]. (b) Druggable hot spots based on druggability simulations using the active form of Src (PDB: 1Y57). Hot spots are shown in yellow spheres. (c) Essential sites predicted using ESSA. Identified essential residues are shown in sticks and transparent spheres. Two new sites are detected by druggability simulations and ESSA, highlighted by red ellipses in the top diagrams (panels b and c).

Druggability simulations performed for the kinase domain highlight multiple sites that may be allosterically targeted (Fig. 2b). Druggability simulations are molecular dynamics (MD) simulations in the presence of explicit water and probe molecules representative of drug-like fragments [26]. The results displayed in Fig. 2b are obtained using as probes benzene, isobutane, imidazole, acetamide, isopropanol and isopropylamine molecules –derived from statistical evaluation of chemical/functional groups most frequently observed in FDA-approved drugs, using the DruGUI [26] and Pharmmaker [28] modules of the ProDy interface [48].

We further identified ‘essential sites’ whose perturbation would elicit cooperative responses in the conformational dynamics of the kinase domain (Fig. 2c), using the Essential Site Scanning Analysis (ESSA) module [33] in ProDy. ESSA is based on modeling the structure as a Gaussian network model (GNM) [49], and scanning all amino acids along the sequence to determine the so-called essential residues whose change in local packing density (potentially induced upon small molecule binding) would elicit a global/cooperative change in the protein’s conformational dynamics. Such sites are proposed to serve as allosteric ligand-binding sites. A recent application to a GPCR demonstrates the utility of utilizing ESSA as a filter for extracting the allosteric sites from amongst multiple druggable sites identified by druggability simulations [30], and designing allosteric modulators that alter function.

Druggability simulations and ESSA applied to kinase domain revealed six sites that were found to be both druggable (shown by yellow spheres in Fig. 2b) and essential (shown by yellow semi-transparent surfaces in Fig 2c): four of them, enclosed in blue ellipses, are known from experiments as ATP-binding (orthosteric), and MYR, PIF, and G-loop (allosteric) sites, and confirmed by simulations to be druggable. Two sites are new (red ellipses in Fig. 2b and c): one close to the PIF site, harbored by the β-sheet at the N lobe, and the other at activation loop. The peptide-substrate site from experiments is also confirmed by druggability simulations to have a propensity to serve as a binding site.

The findings from the druggability simulations and ESSA closely match the results from unbiased ligand binding simulations of Src kinase in the presence of the small molecule ATP-competitive inhibitors PP1 and dasatinib. These simulations showed that PP1 and dasatinib resided in secondary binding sites that matched the known ATP-binding, MYR, PIF-binding sites as well as the G-loop [50].

Elastic network model analysis elucidates the significance of drug-binding sites in controlling the global conformational mechanics of the kinase domain

In order to understand whether the non-ATP-competitive sites play a role in affecting the collective dynamics of the kinase (so that their targeting would impair the function), we performed an elastic network model analysis using the GNM [49], and identified the softest modes of motion accessible to the kinase domain. Each protein of N amino acids has access to N-1 collective modes based on the GNM. The softest modes are those lowest frequency / largest amplitude fluctuations most easily accessible under physiological conditions. Their deployment often facilitates, or assists in, the function of the protein/domain.

The three softest modes of motions deduced from GNM analysis are illustrated by color-coded diagrams in Fig. 3a–c, and the residue displacements along each of these modes’ axes are plotted in panel e-g. The diagrams in panels a-c are color-coded from blue to red, in line with the direction of the movements undergone by the residues along the particular mode. For example, in the softest mode (mode 1; panels a and e), the N- and C-lobes exhibit large, but opposite direction (anticorrelated) movements with respect to each other such that these respective lobes are colored cyan-to-dark blue and orange-to-red, in line with the color bars displayed along the ordinate of panel e. The cleft between the two lobes, on the other hand, which also contains the catalytic residues, serves as a hinge region (and hence colored yellow-to-light green). Residues lying at the hinge region in mode 1 are labeled and highlighted in yellow in panel a; they are distinguished by their minimal (if any) displacement along this mode, and are located at the crossover region between positive and negative displacements in panel e. This mode where the two lobes fluctuate in an anticorrelated way with respect to each other facilitates the transition between the active and inactive forms of the kinase domain, which is characterized by open and closed states, respectively, of the cleft between the N- and C-lobes. Notably, the hinge site partially overlaps with the ATP-binding. This is not a coincidence. It shows that the chemically active (catalytic) site simultaneously also harbors a mechanically key (hinge) site, so as to enable the mechanochemical activity of the kinase domain. Notably, the D₄₀₄FG motif and part of the activation loop also participate in this hinge domain.

Fig. 3. GNM analysis of Src kinase domain.

Residue movements along GNM mode 1 (a), mode 2 (b), and mode 3 (c) are shown in color-coded structures. Panels e-g display the distributions of residue movements when the structure reconfigures along the respective mode’s collective coordinate. Hinge residues in each mode are shown in transparent spheres in the diagrams and highlighted by yellow labels on the diagrams and on the curves. Active form of Src kinase (PDB: 1y57) [43] was used for GNM analysis. (a) Mode 1 enables the opening and closing of cleft between the N- and C-lobes, in line with the transition of the kinase domain between its active and inactive forms. (b) Mode 2 essentially mediates the movements of the G loop; several residues (E412, T440, R460, H492 surrounding (or lying in) the G-loop assume hinge roles. (c) Mode 3 enables the anticorrelated fluctuations between the G-loop and the activation loop revealing an allosteric coupling between those sites. (d) Ribbon diagram color-coded by the size of residue motions (square displacements) driven jointly by modes 1–3. The corresponding square displacements are plotted in panel h. The same panel also displays the motions driven by the individual modes 1, 2 and 3. The vertical yellow shades display the regions that undergo minimal displacements in modes 1–3, thus serving as hinges or anchors that control the global movements of the kinase domain and could serve for allotargeting. The box at the bottom of panel h shows the residue ranges corresponding to substrate-binding (S) and catalytic (C) sites, D₄₀₄FG motif (D), and to the activation loop (A) and G loop (G).

GNM mode 2, on the other hand, places two residues of the G loop (R460 and H492) at the interface between two oppositely moving segments within the C-lobe (Fig. 3b and f), such that the front and back parts of the G-loop undergo concerted fluctuations in opposite directions. GNM mode 3 mediates the anticorrelated motions between the G-loop and the activation loop (Fig. 3c and g). GNM modes 2 and 3 thus suggest that binding of a ligand at the G-loop can dramatically change the conformational dynamics of the protein also affecting the conformation of the activation loop. Panels f and g list the residues located at the hinge regions in these modes.

Fig 3h shows the distribution of residue square-displacements driven by modes 1, 2 and 3, as well as their cumulative contributions (black curve, modes 1–3) and the ribbon diagram in panel d shows the structure color-coded by the size of motion undergone by the individual residues in these three most cooperative modes. Blue regions undergo minimal movements, if any, and serve as hinges or anchors; red regions are the most mobile regions. The light-yellow bands in panels e-h refer to regions that exhibit minimal movements in the cumulative modes 1–3 (panel h). These highly constrained regions include E310 forming the salt-bridge, the catalytic site (residues D386, N391), the D₄₀₄FG motif, residues at the N-terminal end of the activation loop (A408 and R409), and the two ends of the G loop region which serve as anchors while the G loop itself undergoes large scale movements, like the activation loop (see the peaks in panel h black curve for modes 1–3)

Conclusion

With recent advances in computer-aided characterization of structural dynamics using elastic network models, it is increasingly possible to identify novel target sites for allosteric modulation. The latter is particularly useful for designing specific modulators for those targets, like the kinase domain, which contain several conserved sequence and structure motifs shared between family/ subfamily members. Comparison of experimentally verified allotargeting sites and those predicted by computational modeling and simulations - mainly druggability simulations and selection of essential sites from amongst druggable sites using ESSA – clearly demonstrates that in silico studies can be utilized for identifying allosteric sites and designing and developing potent modulators of function. As shown in Fig. 2, experimentally observed allosteric sites are captured by computational models and simulations. Furthermore, insights into the types of collective motions that would be impacted by targeting these allosteric sites (usually located at hinge sites) can be assessed in silico, providing new avenues for specific modulation of enzyme structural dynamics.

It is interesting to observe that the catalytic (ATP-binding) site of the kinase domain also serves as a hinge site in GNM mode 1, mediating the relative movements of the N- and C-lobes. The critical (dynamic) role of ATP binding in controlling the overall dynamics, predicted by a model (GNM) rigorously based on entropic drives, invites attention to the role of conformational entropy maximization in defining the intrinsically accessible global motions. Notably, a recent study reported critical role of ATP-competitive inhibitors in altering conformational entropy [51], further underscoring the importance of entropically driven structural dynamics, in designing inhibitors. The spatial overlap between chemically active (ATP-binding) and mechanically important (hinge) sites also corroborates the mechanochemical behavior of enzymes noted [52] earlier.

We further examined the mechanisms of the softest modes of collective motions, as predicted by the Anisotropic network model (ANM) [53,54]. Supplementary Movie 1 displays the ANM mode 1, which drives the structural changes relevant to the activation/inactivation of the kinase domain. Notably, druggability simulations and ESSA point both to the catalytic site as a druggable and essential site. The former (druggability) property supports the ability of the orthosteric ligands to bind that site, and the latter (essentiality) points to their role in impacting the collective mechanics of the kinase domain. Further examination of other soft modes reveals that ANM mode 3 (in Movie 1) ensures a strong coupling of the reorientation of the activation loop to the collective anticorrelated movements of the N-lobe and C-lobe.

Our in silico studies point to two new allosteric sites, apart from reproducing the allosteric sites known from previous work: one at activation loop, and the other is near the G loop (see ANM modes 1–3 in Movie 1). The activation loop itself thus serves as a binding site for allosteric modulators, to induce functional changes in conformation. Notably, recent NMR Carr-Purcell-Meiboom-Gill relaxation dispersion experiments invited attention to a coupling between the motions of the phosphorylated activation loop of ERK2 and the global motions of the kinase domain facilitate nucleotide binding to the catalytic domain [55]. The involvement of the activation loop in this type of global change in structure is consistent with cooperative dynamics in the softest (most easily accessible) modes predicted by the GNM.

According to GNM modes 2 and 3, the G-loop is another allosteric site. In GNM mode 2, hinge residues lying at G-loop site controlling the large scale movements of the G loop, and mode 3 mediates the coupled motions between the G-loop and the activation loop sites, meaning that ligand binding at the G-loop can potentially alter the conformation of the activation loop.

The emerging picture about the kinase domain structural dynamics is that of a finely tuned module, with a typical two-lobe architecture and specific loop regions, which enables the predisposition of protein kinases to integrate various allosteric stimuli and relay them to the orthosteric site which affects substrate phosphorylation [51]. Vice versa, ligand binding to the orthosteric side can affect the dynamics of allosteric sites and potentially alter substrate binding or interaction with regulatory domains [56–59]. Similarly, resistance mutations may not only reduce the binding affinity of kinase inhibitors, but also the residence time – a key predictor for drug efficacy [60–63]. Consequently, combination therapy with allosteric and orthosteric inhibitors requires compatible inhibitor types to maximize synergy for potency and increased residence time [59,64–67].

Data availability

Data will be made available on request.