https://orcid.org/0000-0002-5256-9456
Abstract
Cyclin-dependent kinase (Cdk) proteins play crucial roles in the cell cycle progression and are thus attractive drug targets for therapy against such aberrant cell cycle processes as cancer. Since most of the available Cdk inhibitors target the highly conserved catalytic ATP pocket and their lack of specificity often lead to side effects, it is imperative to identify and characterize less conserved non-catalytic pockets capable of interfering with the kinase activity allosterically. However, a systematic analysis of these allosteric druggable pockets is still in its infancy. Here, we summarize the existing Cdk pockets and their selectivity. Then, we outline a network-based pocket prediction approach (NetPocket) and illustrate its utility for systematically identifying the allosteric druggable pockets with case studies. Finally, we discuss potential future directions and their challenges.
Overview of the Cdks
Cyclin-dependent kinases (Cdks) regulate cell cycle phases by interacting with their cyclin partners [1]. Cdks play crucial roles mainly in the following processes: mitogenic signals, oncogene-induced or replicative senescence, growth-inhibitory signals, transcriptional functions and DNA damage (Figure 1) [1–3]. Cdks fall into two categories based on their functions. The first group comprises Cdk1, Cdk2 and Cdk4/6, which directly drive the cell cycle [4, 5]. For example, Cdk1 is activated by Cyclin A at the end of the interphase to facilitate the onset of mitosis [6, 7]. Cdk2 is essential for the G1/S transition upon binding with Cyclin E [8, 9], and Cdk4/6 is also responsible for cell proliferation [10, 11]. Any aberrant activation of these Cdks is frequently associated with human cancers [2, 12–14]. The second group consists of Cdk7, Cdk8, Cdk9, Cdk11 and Cdk20 [4], which play essential roles in transcription. For instance, Cdk9 and Cyclin T1 form the positive transcription elongation factor b (P-TEFb), triggering the transition of promoter-proximal Pol II from pause to elongation [15–17]. In the context of virus infection, this crucial elongating function is thus hijacked by the human immunodeficiency virus (HIV) for rapid virus replication [18, 19]. Therefore, Cdks are considered attractive drug targets, especially for stopping cancer and virus infection.
Significant research progress related to Cdks. The highlighted boxes represent the milestones for the Cdk study. The bottom left is the FDA-approved Cdk4/6 inhibitors.
Cdks (EC 2.7.11.22) belong to the CMGC kinase group and consist of 20 Cdks from Cdk1 to Cdk20 [20]. The available crystal structures of Cdk proteins contain ~300 amino acids [18, 21, 22]. We calculated the sequence conservation for all Cdks. The continuous conservation scores are divided into a discrete scale of nine grades, with grades 1–3 for the most variable positions and 7–9 for the most conserved positions. For example, Cdk2 has 30.2% variable residues (conservation-grade 1–3), 29.5% medium conserved residues (conservation-grade 4–6) and 40.3% conserved residues (conservation-grade 7–9), respectively (Figure S1, see Supplementary Data available online at https://academic.oup.com/bib). Structurally, Cdks have a two-lobed structure: the N-terminal lobe with β-sheets and the C-terminal lobe rich in α-helices [23]. The active site is in a sandwiched conformation. The N-lobe contains a glycine-rich inhibitory element (G-loop) and a unique major helix (C-helix) [24]. The C-lobe contains the activation segment, spans from the DFG motif to the APE motif and includes the phosphorylation site in the T-loop [4]. The Cdks are highly similar in sequence and structure. The adjacent subfamilies’ root-mean-square deviation (RMSD) values are around 1 Å (Figure 2). Thus, these detailed analyses demonstrate that the catalytic inhibitor may affect multiple Cdk proteins simultaneously and lead to side effects [1–3].
The evolutionary and similarity analysis of the Cdk family. The representative crystal structures are superimposed on the right. The structural differences are less than 1 Å.
Many experimental methods have been brought to determine the structure determination of Cdk/inhibitor complexes. For example, X-ray diffraction (XRD), nuclear magnetic resonance (NMR) and cryogenic electron microscopy (cryo-EM) can solve the Cdk/inhibitor structure and provide information on the interface interaction. However, additional experiments, such as electrophoretic mobility shift assays or surface plasmon resonance (SPR), are still required for binding affinity quantification. Therefore, the experimental methods for the druggable pocket detection are expensive and time-consuming. Some existing computational methods can predict the pocket of the protein/inhibitor complex: (i) the template-based methods (FINDSITE [25]) are difficult to predict the pocket when the template is not available; (ii) the probe-based (FTSite [26]) or deep learning-based (DeepPocket [27]) methods highly dependent on the choices of probes and energy functions. In addition, the existing methods are challenging in probing the minor allosteric dynamical consequences for druggable pocket identification. Here, we outline a network-based pocket prediction approach (NetPocket) and illustrate its utility for systematically identifying the allosteric druggable pockets with case studies.
The selective pockets and inhibitors
The available Cdk/inhibitor complex structures revealed two distinctive strategies targeting Cdk proteins: (i) ATP-competitive inhibitors targeting the catalytic ATP-binding pocket (type I and II) and (ii) the allosteric inhibitors targeting allosteric pockets (type III) or pockets located at the interface for Cdk partners (type IV) [28]. In addition, the sequence and structural analyses of the pockets can provide explanations of the inhibitor selectivity and side effects. For example, the sequences of Cdk2 and Cdk9 were derived from the RCSB Protein Data Bank (PDB ID: 1FIN and 6CYT). The homology sequences and continuous conservation scores (Figure 3A and B) were calculated using the empirical Bayesian model by the ConSurf server [29–31]. Then, the WebLogo plots were generated to show the sequence conservation [32]. The overall height of the sequences in WebLogo indicates the frequency and conservation at the corresponding position. As shown in Figure 3C–F, the sequences of the ATP-pocket are highly similar in Cdk2 and Cdk9, while the sequences of the TL-pocket are highly variable.
The sequence analysis of the ATP-binding pockets and non-catalytic TL-pockets on Cdk2 and Cdk9. The ATP-binding pockets and TL-pockets of (A) Cdk2 and (B) Cdk9 are circled in red color. The residues are labeled with different colors according to the sequence conservation. The multiple sequence alignment for (C) Cdk2 and (D) Cdk9 shows that the ATP-binding pockets are highly conserved. The residues are labeled with different colors according to the sequence conservation. The multiple sequence alignment for (E) Cdk2 and (F) Cdk9 shows that both sequence and structure of the TL-pockets are highly variable.
The catalytic pocket and inhibitors
The ATP-binding pocket is located between the N-terminal lobe and C-terminal lobes. The catalytic site is regulated by the flexible activation loop (A-loop, 20–30 amino acids) [33]. Structurally, the catalytic pocket is larger and more rigid than the non-catalytic pockets [28]. The type I inhibitors directly interact with the ATP binding sites of the active form of the Cdk, which is in the DFG-in state. Type II inhibitors interact with a DFG-out and inactive Cdk conformation [28]. Currently, there are several developed inhibitors under clinical trials phase I–IV.
As shown in Table 1, the ATP competitive inhibitors target several Cdk proteins simultaneously, leading to side effects or even toxicity. We analyzed the sequence conservation of the ATP-binding pockets between Cdk2 and Cdk9 (Figure 3C and D). The continuous conservation scores are divided into a discrete scale of 9 grades for visualization purposes. Grade 1 contains the most variable positions colored in turquoise, and grade 9 contains the most conserved positions colored in maroon. The conservation-grade colors are projected onto the Cdk tertiary structure [34]. The homology sequence analysis shows that the ATP-binding pockets are highly conserved at the sequence level [35]. Structurally, the catalytic ATP-binding pockets between Cdk2 and Cdk9 are similar, with a small RMSD of 0.42 Å. These characteristics of ATP-pocket are likely to account for the side effects of catalytic inhibitors. For instance, Flavopiridol was developed to target Cdk9 for acquired immune deficiency syndrome (AIDS) [36, 37]. However, previous experiments found that Flavopiridol would also inhibit the normal functions of Cdk1, Cdk2, Cdk4 and Cdk6 [36, 38, 39]. Thus, Flavopiridol induces apoptosis in several mouse tissues, leading to organ atrophy [40, 41]. The second-generation pan-Cdk inhibitor Dinaciclib was developed to solve the above issues [42]. The results showed that Dinaciclib has selectivity and higher potency in inhibiting RB phosphorylation [43]. This inhibitor can block the proliferation of ovarian and pancreatic tumor cells [44, 45]. Recently, FDA approved three ATP-competitive Cdk inhibitors: Palbociclib, Ribociclib and Abemaciclib. They are selective Cdk 4/6 inhibitors, which inhibit Cdk4 and Cdk6 by binding to the catalytic ATP-binding pocket [46]. Palbociclib and Ribociclib have similar Cdk4/6 inhibition potencies with IC50 of 11/15 nM (for Palbociclib) [47] and 10/39 nM (for Ribociclib) [48]. Abemaciclib shows a lower IC50 of 2 nM for Cdk4 and 10 nM for Cdk6 [49]. However, Abemaciclib has an additional activity against Cdk9, a crucial regulating unit in transcriptional events in embryogenesis and cell proliferation [50]. Therefore, these three approved inhibitors also exhibit toxicity [51–53].
Inhibitors for the ATP-binding pocket
| Inhibitor | Primary target(s) | Clinical trials | Reference |
|---|---|---|---|
| AG-024322 | Cdk1, Cdk2, Cdk4 | Phase I | [107] |
| AT7519 | Cdk1, Cdk2, Cdk4, Cdk5 | Phase I/II | [108] |
| R547 | Cdk1, Cdk2, Cdk4, Cdk7 | Phase I | [109] |
| Flavopiridol | Cdk1, Cdk2, Cdk4, Cdk6, Cdk9 | Phase II | [41] |
| Dinaciclib | Cdk1, Cdk2, Cdk5 Cdk9 | Phase I/II | [43] |
| Voruciclib | Cdk4/6 | Phase I | [110] |
| Abemaciclib | Cdk4/6 | Phase IV | [49] |
| Palbociclib | Cdk4/6 | Phase IV | [111] |
| Ribociclib | Cdk4/6 | Phase IV | [48] |
| Inhibitor | Primary target(s) | Clinical trials | Reference |
|---|---|---|---|
| AG-024322 | Cdk1, Cdk2, Cdk4 | Phase I | [107] |
| AT7519 | Cdk1, Cdk2, Cdk4, Cdk5 | Phase I/II | [108] |
| R547 | Cdk1, Cdk2, Cdk4, Cdk7 | Phase I | [109] |
| Flavopiridol | Cdk1, Cdk2, Cdk4, Cdk6, Cdk9 | Phase II | [41] |
| Dinaciclib | Cdk1, Cdk2, Cdk5 Cdk9 | Phase I/II | [43] |
| Voruciclib | Cdk4/6 | Phase I | [110] |
| Abemaciclib | Cdk4/6 | Phase IV | [49] |
| Palbociclib | Cdk4/6 | Phase IV | [111] |
| Ribociclib | Cdk4/6 | Phase IV | [48] |
Inhibitors for the ATP-binding pocket
| Inhibitor | Primary target(s) | Clinical trials | Reference |
|---|---|---|---|
| AG-024322 | Cdk1, Cdk2, Cdk4 | Phase I | [107] |
| AT7519 | Cdk1, Cdk2, Cdk4, Cdk5 | Phase I/II | [108] |
| R547 | Cdk1, Cdk2, Cdk4, Cdk7 | Phase I | [109] |
| Flavopiridol | Cdk1, Cdk2, Cdk4, Cdk6, Cdk9 | Phase II | [41] |
| Dinaciclib | Cdk1, Cdk2, Cdk5 Cdk9 | Phase I/II | [43] |
| Voruciclib | Cdk4/6 | Phase I | [110] |
| Abemaciclib | Cdk4/6 | Phase IV | [49] |
| Palbociclib | Cdk4/6 | Phase IV | [111] |
| Ribociclib | Cdk4/6 | Phase IV | [48] |
| Inhibitor | Primary target(s) | Clinical trials | Reference |
|---|---|---|---|
| AG-024322 | Cdk1, Cdk2, Cdk4 | Phase I | [107] |
| AT7519 | Cdk1, Cdk2, Cdk4, Cdk5 | Phase I/II | [108] |
| R547 | Cdk1, Cdk2, Cdk4, Cdk7 | Phase I | [109] |
| Flavopiridol | Cdk1, Cdk2, Cdk4, Cdk6, Cdk9 | Phase II | [41] |
| Dinaciclib | Cdk1, Cdk2, Cdk5 Cdk9 | Phase I/II | [43] |
| Voruciclib | Cdk4/6 | Phase I | [110] |
| Abemaciclib | Cdk4/6 | Phase IV | [49] |
| Palbociclib | Cdk4/6 | Phase IV | [111] |
| Ribociclib | Cdk4/6 | Phase IV | [48] |
The non-catalytic pocket and inhibitors
The type III inhibitor targets allosteric pockets and usually leads to structural changes in the ATP-binding pocket. The type IV inhibitor targets the pocket between Cdk and their partners. The previous studies showed that 85% of the 1717 identified pockets are non-catalytic in the HKpocket database [54]. Most of these pockets are highly flexible and smaller than the ATP-binding pocket. Currently, there are several developed non-catalytic inhibitors in the pre-clinical phase.
As shown in Table 2, previous research identified two non-catalytic pockets on Cdk proteins. Betzi et al. reported one potential allosteric inhibitor, ANS, targeting a pocket adjacent to the C-helix of Cdk2. Their X-ray results show that the ANS can induce significant structural changes in Cdk2 and further regulate the interactions between Cdk2 and Cyclin A [55]. As shown in Figure S2 (see Supplementary Data available online at https://academic.oup.com/bib), the binding site of ANS is located approximately midway between the ATP-binding site and the C-helix but away from the ATP site, and two distinct ANS molecules are bound adjacent to each other. Zhao et al. reported that the inhibitor F07#13 could break up the Cdk9/Cyclin T1 interface by targeting a T-loop pocket (TL-pocket) [56, 57]. The TL-pocket is formed in part by the T-loop of Cdks, which is usually located under the ATP-binding pocket. The TL-pocket usually contains 15–20 residues, for example, residue 150–162 in Cdk2 and residue 170–185 in Cdk9 [57, 58]. The multiple sequence alignment shows that the sequences vary in Cdks (Figure 3E and F). Structurally, the volume, surface and depth differences in TL-pocket across Cdks are significantly larger than those in the ATP-binding pocket. The experiment in vitro shows that phosphorylation of both the CTD domain of RNA Pol II and histone H1 are inhibited by F07#13 [57]. In the Cdk2 case study, the conformation of the T-loop region shows sharp perturbation upon peptide binding. The simulation study of Cdk2 shows that its glycine-rich loop moves away from the ATP binding pocket [59]. As shown in Figure S3 (see Supplementary Data available online at https://academic.oup.com/bib), the T-loop of Cdk2 is very flexible in the unbound state (PDB ID: 1E1X). The T-loop becomes more rigid when Cdk2 binds to Cyclin protein [60]. Then, the Thr160 in the T-loop is phosphorylated for functions. The T-loop pocket is heavily involved in the Cdk2/Cyclin complex formation and function [60]. The allosteric binding pockets are far less conserved than the ATP-binding pocket, accommodating more specific inhibitors against different Cdks.
Generally, the non-catalytic pocket represents a variable sequence and flexible structure. Therefore, we compared the geometry information of the ATP-binding pocket and TL-pocket. As shown in Figure 4 and Table 3, the geometry property values (volume, surface and depth) of the ATP-binding pockets in Cdk2 and Cdk9 proteins are very similar. The ATP-binding pocket volume of Cdk9 (651.2 Å3) is almost the same as the Cdk2 (683.0 Å3). The rmsd of ATP-binding pockets between Cdk2 and Cdk9 is 0.42 Å (calculated by Chimera) [61]. The ATP-competitive inhibitor permanently bounds to several Cdks and explains that Flavopiridol can target Cdk1, Cdk2, Cdk4, Cdk6 and Cdk9 for side effects. In contrast, the geometry property values of the TL-pocket in Cdk2 are significantly different from TL-pocket in Cdk9. The volume, surface and depth differences are 12.9, 28.6 and 35.5%. The results suggest that the non-catalytic pocket is more selective for inhibitors to reduce the side effects.
Geometrical comparison analysis of the pockets on Cdk2 and Cdk9. The geometrical values of the ATP-pockets (colored in orange) on (A) Cdk2 and (B) Cdk9 are similar. However, the geometrical characteristic of the TL-pockets (colored in green) on Cdk2 and Cdk9 shows significant differences (C).
The comparison of ATP-binding and TL-pocket geometric values between Cdk2 and Cdk9
| Volume/ Å3 | Surface/ Å2 | Depth/ Å | |
|---|---|---|---|
| ATP-pocket (Cdk2) | 651.3 | 706.4 | 16.2 |
| ATP-pocket (Cdk9) | 683.0 | 652.8 | 18.3 |
| Difference | 4.8% | 7.9% | 12.2% |
| TL-pocket (Cdk2) | 330.7 | 593.6 | 17.6 |
| TL-pocket (Cdk9) | 376.2 | 445.2 | 12.3 |
| Difference | 12.9% | 28.6% | 35.5% |
| Volume/ Å3 | Surface/ Å2 | Depth/ Å | |
|---|---|---|---|
| ATP-pocket (Cdk2) | 651.3 | 706.4 | 16.2 |
| ATP-pocket (Cdk9) | 683.0 | 652.8 | 18.3 |
| Difference | 4.8% | 7.9% | 12.2% |
| TL-pocket (Cdk2) | 330.7 | 593.6 | 17.6 |
| TL-pocket (Cdk9) | 376.2 | 445.2 | 12.3 |
| Difference | 12.9% | 28.6% | 35.5% |
The comparison of ATP-binding and TL-pocket geometric values between Cdk2 and Cdk9
| Volume/ Å3 | Surface/ Å2 | Depth/ Å | |
|---|---|---|---|
| ATP-pocket (Cdk2) | 651.3 | 706.4 | 16.2 |
| ATP-pocket (Cdk9) | 683.0 | 652.8 | 18.3 |
| Difference | 4.8% | 7.9% | 12.2% |
| TL-pocket (Cdk2) | 330.7 | 593.6 | 17.6 |
| TL-pocket (Cdk9) | 376.2 | 445.2 | 12.3 |
| Difference | 12.9% | 28.6% | 35.5% |
| Volume/ Å3 | Surface/ Å2 | Depth/ Å | |
|---|---|---|---|
| ATP-pocket (Cdk2) | 651.3 | 706.4 | 16.2 |
| ATP-pocket (Cdk9) | 683.0 | 652.8 | 18.3 |
| Difference | 4.8% | 7.9% | 12.2% |
| TL-pocket (Cdk2) | 330.7 | 593.6 | 17.6 |
| TL-pocket (Cdk9) | 376.2 | 445.2 | 12.3 |
| Difference | 12.9% | 28.6% | 35.5% |
Allosteric inhibitor pocket prediction methods
Identifying inhibitor targets plays a critical role in understanding the disease mechanisms and drug development. The inhibitor target identification methods can be divided into the biological experiment and computational prediction categories. In the experiments, western blot, ITC and SPR can detect the target/inhibitor binding. However, the target/inhibitor interaction details need to be determined by X-ray, NMR or cryo-EM. For the theoretical methods, molecular docking and molecular dynamics (MD) simulations can predict the target/inhibitor interface and provide the specific binding poses. However, there are two additional challenges for allosteric pockets: finding the pockets that are amenable for small molecule docking and predicting the consequence of the binding, i.e. how the allosteric pocket is linked to the catalytic function. Therefore, the allosteric inhibitor target prediction methods are urgently needed.
Network-based pocket prediction and inhibitor screen
The previous research of the Zhao group showed that the NetPocket (Network-based Pocket) approach is efficient and effective for predicting allosteric inhibitor pockets [54, 57, 62–64]. The NetPocket can be described in four steps [65–67] (Figure 5).
The workflow of the NetPocket. Firstly, DoGSiteScorer is employed to analyze the topology information of potential pockets. Then, the dynamical network analysis is adopted to detect the correlations between the allosteric pockets and the functional sites. Next, the contact map and MM-PBSA analyze the inhibitor binding affinity. Finally, the SMD is used to verify the binding.
The first step is to identify the potential pockets using the rolling probe method [68, 69]. The pocket is detected by calculating the rolling translational degrees of freedom. The center of the probing sphere is recorded if the sphere contacts more than two atoms on the molecule. These center positions describe the detected pocket boundary of the protein under consideration. Next, the geometrical information of pockets was calculated using the discrete volume method [26].
The interaction energy decomposition contributions can be used to discover the key residues that play crucial roles in binding the inhibitor to Cdk. Therefore, mutations on these critical residues on the pocket can be used to verify the allosteric mechanism, for example, by using the final step described below.
Case studies
The NetPocket has been successfully applied in several inhibitor screen studies. For example, the Cdk/Cyclin complex can promote progression from the G1 phase into the S phase (Cdk2/Cyclin E) and play roles in the S phase (Cdk/Cyclin A) for DNA synthesis. In addition, Cdk2 is associated with mammary cancer formation in mouse cancer models [86, 87]. To inhibit the activity of Cdk2, Chen et al. [58] performed the NetPocket to analyze the dynamical correlations between each potentially allosteric pocket and Cdk2/Cyclin interface. The dynamical correlations showed that the non-catalytic TL-pocket is highly correlated with the Cdk2/Cyclin interface. Thus, Chen et al. designed several peptides targeting the TL-pocket to break the Cdk2/Cyclin interface allosterically. The dynamical network analysis revealed that the Cdk2/Cyclin interface was significantly weakened upon the peptide binding. Furthermore, the SPR and western blot experiments showed that two of these peptides, DAALT and YAALQ, break up the CDK2/Cyclin complex partially and diminish its kinase activity in vitro.
The HIV is a retrovirus that progressively attacks the human immune system [88, 89]. The HIV viral protein Tat recruits the positive transcription elongation factor b (p-TEFb) protein complex to enhance its viral transcription [90, 91]. As a result, Tat binds to both Cdk9 and Cyclin T1 and thus creates a second pathway between Cdk9 and Cyclin T1 [92–94]. Zhao et al. [57] applied the NetPocket to identify the TL-pocket on Cdk9 as an allosteric pocket in the presence of Tat since the TL-pocket becomes correlated with the Cdk9/Cyclin interface via this second pathway. The dynamical network simulations indicate a noticeable interface weakening upon binding the inhibitor F07#13 to the TL-pocket. Indeed, the immunoprecipitation experiment verified that the inhibitor F07#13 decreased the kinase activity of Cdk9 in the HIV-infected cells. The TL-pocket affects the binding interface of Cdk9 and Cyclin T1 in the infected cells. The F07#13 is thus a type IV inhibitor that breaks up the interface and inhibits the complex formation.
Moreover, Wang et al. [65] systematically applied NetPocket to characterize the non-catalytic pockets for the structurally well-covered human kinome. Kinase proteins have been intensively investigated as inhibitor targets for decades. However, most available kinase inhibitors target the ATP-binding pocket resulting in side effects. It is thus highly desirable to improve selectivity by inhibiting kinase activity allosterically. In their study (54), the 29 group-specific non-catalytic (GSNC) pockets on 168 representative kinase proteins were first clustered into seven groups based on their location and shape distances. Then, the dynamical network analysis was performed to detect the subtle residue–residue interaction changes to elucidate the complex allosteric consequences. Compared with the experiments, 14 group-specific allosteric pockets are promising targets that may benefit kinase drug design. The experiment performed by Comess et al. discovered a small non-catalytic pocket in JNK1α1 that is targeted by inhibitor A-82118 (PDB ID: 3O2M). This pocket is located at the same position as pocket p2 of CLK1 kinase in the CMGC group [95]. Moreover, the allosteric inhibitors 4-(carbamoylamino)-1-(7-ethoxynaphthalen-1-yl)-1H-pyrazole-3-carboxamide (PDB ID: 4 M12) were developed to target a non-catalytic pocket of ITK kinase, which is located at the same position as pocket p6 of JAK1 in the TK group [96]. These results demonstrate that the predicted GSNC pockets may indeed act as allosteric sites for inhibitor binding.
Future directions
The NetPocket approach has advances over existing methods (Table 4). (i) The existing methods do not consider the pocket topology information which is crucial for inhibitor design. For example, Lai et al. used a GNM-based approach to discover the motions of known allosteric sites and the related sites [97]. Panjkovich and Daura [98] used an NMA-based approach to estimate structural flexibility. These two methods are fast and applicable but without the pocket topology calculations. (ii) The existing methods are challenging in probing the allosteric dynamical consequences, such as pocket flexibility and motion correlations. For example, Huang et al. [99] developed an SVM-based pocket detection algorithm. Goncearenco et al. [100] adopted the scoring function to predict potential ligand-binding sites. However, these two methods could not consider the allosteric dynamical consequences. The NetPocket performed the dynamical network correlations with integrated the MM-PBSA calculations to identify the little allosteric dynamical consequences. (iii) The NetPocket can probe the selectivity of allosteric pockets, which are critical for side effects. For example, NetPocket has successfully identified the selectivity TL-pocket using the NetPocket approach. The F07#13 is an allosteric inhibitor designed to target the TL-pocket in the Cdk9 protein. To verify the selectivity, we predicted the Cdk/F07#13 complex structures and highlighted the top3 binding conformations using HDOCK [101] (Figure 6). All the top3 conformations bind to the TL-pocket in Cdk9/F07#13 predictions [57]. However, the other top3 conformations show that F07#13 attaches to various pockets in Cdk2, Cdk4 and Cdk6 prediction cases.
The Cdk/F07#13 complex structure prediction using HDOCK. (A) shows the overlapping structures between the Cdks. The (B–E) are the top3 F07#13 binding conformations in Cdk2, Cdk4, Cdk6 and Cdk9 predictions. All the top3 conformations bind to the TL-pocket in Cdk9/F07#13 predictions. However, the other top3 conformations show that F07#13 attaches to various pockets in Cdk2, Cdk4 and Cdk6 prediction cases.
The framework of allosteric pocket prediction and inhibitor screening with deep learning. The first network consists of an input layer, time series (1 − τ), hidden layer and output layer. The input layer contains the atom, structure, topology and binding dynamics information, and the output is the correlation between the allosteric pocket to the functional domain. The input layer of the second network is the output by the first network and the information of the inhibitor, and the output is the correlation and matching score of pocket and inhibitor.
Geometrical comparison analysis of the non-catalytic pockets on the CMGC family. (A) The non-catalytic pockets are highlighted in different colors, which correspond to the (B) volume, (C) surface and (D) depth, respectively. The topology information of the non-catalytic pockets in the volume, surface or pocket depth shows a significant difference compared with ATP-pocket in the CMGC family.
The NetPocket utilizes both dynamics network correlation to identify the small structural changes and steered molecular dynamics to amplify the consequences. Previous research shows an efficient and precise prediction of the allosteric druggable pockets. However, NetPocket remains two limitations. First, the choice of force fields for MD depends on the complex structure. Therefore, short MD simulations need to be performed to test the force fields separately since the appropriate force fields may provide more accurate performance. Second, the choice of the vector and spring constant (k) of a steering force for steered molecular dynamics depends on the experience. The different options for the force parameter may lead to different results that deviate from the crystal structure. Therefore, it is crucial to select the proper steered force through multiple tests. Therefore, one intuitive and easy-to-use deep learning-based approach for allosteric druggable pocket prediction needs to be developed [102]. AlphaFold2 and ARES have been successfully used for protein and RNA tertiary structure prediction [103, 104]. Their methods utilize a deep neural network to predict the RMSD from the unknown structure by each atom’s 3D coordinates and chemical element types. Unlike Townshend Raphael et al. [103] using the static structural information, the simulation trajectories are needed for dynamical correlations identification. Besides, the atomic information, such as bond and angle, the topology of the local domain and the global structural characteristic, need to be considered. As shown in Figure 7, the first deep network can be used to predict the conformational characteristics of the allosteric pocket and the impact on the global structure using MD simulations. Then, the second deep network screens for appropriate inhibitors with ranking scores. This intuitive approach may facilitate the allosteric druggable pocket prediction and accelerate the drug design development.
The topological characteristic of the catalytic pocket is usually highly similar (Figure 4). But the allosteric druggable pockets are widely varied. The non-catalytic pockets are shallow and small compared with the ATP-binding pocket in the CMGC (Figure 8), AGC (Figure S4, see Supplementary Data available online at https://academic.oup.com/bib), CAMK (Figure S5, see Supplementary Data available online at https://academic.oup.com/bib), CK1 (Figure S6, see Supplementary Data available online at https://academic.oup.com/bib), TK (Figure S7, see Supplementary Data available online at https://academic.oup.com/bib) and TKL (Figure S8, see Supplementary Data available online at https://academic.oup.com/bib) family calculations. Therefore, it is more challenging to screen for or design the allosteric inhibitors of high binding affinity than the catalytic ones. An alternative and perhaps more practical approach would be to develop allosteric inhibitors simultaneously targeting two or more allosteric pockets. For example, the HIV viral protein Tat recruits the host P-TEFb onto the nascent HIV viral transactivation response element (TAR) RNA to overcome the elongation pause for active transcription [105]. Ning et al. [67] probed the structural ensembles and binding dynamics of various interfaces in the Tat/TAR/P-TEFb complex. The results show that the TL-pocket inhibitor F07#13 dissociated the Tat from the P-TEFb complex. Thus, Tat could not recruit the P-TEFb to initiate transcription upon TL-pocket inhibitor binding. On the other hand, the inhibitor JB181 binds to the Tat/TAR interface [106]. This inhibitor can break up the Tat/TAR interactions and prevent TAR from pulling down the Tat from the P-TEFb protein. With both inhibitors applied simultaneously, the simulations indicate a far more complete removal of Tat from the complex and thus more efficient disintegration of the complex (Figure S9, see Supplementary Data available online at https://academic.oup.com/bib). Allosteric druggable pocket prediction research is promising but requires more effort to reduce its limitations for rapid growth.
The inhibitors that target the highly conserved catalytic ATP pocket often interact with multiple Cdk proteins simultaneously leading to side effects. Previous research suggests that non-catalytic pockets are more diverse and thus present a class of promising targets to develop inhibitors of fewer side effects.
The NetPocket approach probes the dynamical network correlations via MM-PBSA MD simulations to detect the subtle allosteric dynamics and its consequences.
The NetPocket approach quantifies the selectivity of allosteric pockets, which are critical for reducing side effects.
Authors’ contributions
S.N. performed the analysis and wrote the paper; H.W. helped with the data collection; C.Z. helped with the discussion; Y.Z. supervised the overall study and edited the manuscript.
Funding
National Natural Science Foundation of China (grant no. 12175081); Fundamental Research Funds for the Central Universities (grant no. CCNU22QN004); Excellent doctorial dissertation cultivation grant from Central China Normal University (grant no. 2022YBZZ043).
Data Availability Statement
The data underlying this article are available in the article and in its online supplementary material.
Abbreviations
Cdk, Cyclin-dependent kinase; P-TEFb, Positive transcription elongation factor b; Pol II, Polymerase II; DFT-in/out, Asp (D)–Phe (F)–Gly (G) active and inactive state; TL-pocket, T-loop pocket; CTD/NTD domain, C-terminal/N-terminal domain; ASMD, Adaptive steered molecular dynamics; MD, Molecular dynamics; MM-GBSA, Molecular mechanics-generalized Born surface area
Author Biographies
Shangbo Ning is working toward a Ph.D. in the College of Physical Science and Technology at Central China Normal University, Wuhan, China. His current research interests include biophysics, bioinformatics, and drug design.
Huiwen Wang received the Ph.D. degree in the College of Physical Science and Technology at Central China Normal University, Wuhan, China, in 2021. She is currently an associate professor at the School of Physics and Engineering at Henan University of Science and Technology, Luoyang, China. Her current research interests include biophysics and deep learning.
Chen Zeng is a professor in the Department of Physics at George Washington University, Washington DC, USA. His main research interests include biophysics, structural biology, and drug design.
Yunjie Zhao is a professor in the College of Physical Science and Technology at Central China Normal University, Wuhan, China. His research interests focus on fundamental principles underlying molecular structure and function; encoding the principles for biomolecule design related to human diseases; developing the computational tools to address challenges in health and technology.
