Comprehensive structural and functional characterization of the human kinome by protein structure modeling and ligand virtual screening

Abstract

The growing interest in the identification of kinase inhibitors, promising therapeutics in the treatment of many diseases, has created a demand for the structural characterization of the entire human kinome. At the outset of the drug development process, the lead-finding stage, approaches that enrich the screening library with bioactive compounds are needed. Here, protein structure-based methods can play an important role, but despite structural genomics efforts, it is unlikely that the three-dimensional structures of the entire kinome will be available soon. Therefore, at the proteome level, structure-based approaches must rely on predicted models, with a key issue being their utility in virtual ligand screening. In this study, we employ the recently developed FINDSITE/Q-Dock Ligand Homology Modeling approach, which is well suited for proteome-scale applications using predicted structures, to provide extensive structural and functional characterization of the human kinome. Specifically, we construct structure models for the human kinome; these are subsequently subject to virtual screening against a library of more than 2 million compounds. To rank the compounds, we employ a hierarchical approach that combines ligand- and structure-based filters. Modeling accuracy is carefully validated using available experimental data with particularly encouraging results found for the ability to identify, without prior knowledge, specific kinase inhibitors. More generally, the modeling procedure results in a large number of predicted molecular interactions between kinases and small ligands that should be of practical use in the development of novel inhibitors. The dataset is freely available to the academic community a user-friendly web interface at http://cssb.biology.gatech.edu/kinomelhm/as well as the ZINC website (http://zinc.docking.org/applications/2010Apr/Brylinski-2010.tar.gz).

1. INTRODUCTION

One of the largest enzyme families, the protein kinase family, comprises about ~2% of the human proteome ¹. Each member of this family contains a highly conserved kinase catalytic domain responsible for the reversible phosphorylation of protein substrates, a major regulatory process in both prokaryotic and eukaryotic organisms ^{2, 3}. The transfer of the γ-phosphate of ATP to serine, threonine and tyrosine residues in many enzymes and receptors turns them on and off; thus, the dysfunction of kinase activity is implicated in various pathological conditions. The regulation of kinase activity has been recognized by the pharmaceutical industry as an important therapeutic strategy in the treatment of many diseases including cancer, Alzheimer’s disease, diabetes, inflammation, multiple sclerosis and cardiovascular disease ^4–8. Currently, an estimated one-third of drug discovery programs focus on protein kinases ⁹, with already approved drugs such as imatinib ¹⁰ (Gleevec, Novartis), gefitinib ¹¹ (Iressa, AstraZeneca), lapatinib ¹² (Tykerb/Tyverb, GlaxoSmithKline) or sunitinib ¹³ (Sutent, Pfizer). These are just a few of the more than a hundred successfully developed compounds with kinase inhibition as their mode of action ¹⁴.

To speed up the development of new biopharmaceuticals, computational techniques for the identification of lead compounds are widely used ¹⁵. In particular, virtual screening, a technique that shows great promise for lead discovery, is becoming an integral part of modern drug design pipelines ^{16, 17}. Due to advances in computer technology resulting in constantly increasing computational power, virtual libraries comprising millions of compounds can be rapidly evaluated in silico prior to experimental screens and at the fraction of the cost. Virtual screening approaches, historically divided into ligand- and structure-based algorithms ¹⁸, prioritize drug candidates by estimating the probability of binding to the target receptor. Among many methods developed to date, docking-based techniques are valuable tools for lead identification ¹⁹. These algorithms rank compounds by predicting the binding mode for a query molecule in the binding pocket of the target protein ^20–22; this is followed by the prediction of binding affinity from molecular interactions ^23–25. Recent successful applications of structure-based virtual screening to kinase targets include the identification of potent inhibitors for death-associated protein kinases (DAPKs) ²⁶, protein kinase B (PKB/AKT) ²⁷, Janus kinase 2 (JAK2) ²⁸, Met tyrosine kinase (RTK Met) ²⁹ and Aurora kinase A (AurA) ³⁰.

Notwithstanding the practical value of virtual screening by ligand docking for lead identification, there are significant flaws in current methods. Most salient is the fact that the predicted binding affinity is strongly correlated with the molecular weight of the ligand, independent of whether or not the ligand really binds to its target ^{31, 32}. Furthermore, to achieve satisfactory performance, many commonly used docking algorithms require the X-ray structure of their receptor target, preferably in the ligand-bound conformational state ³³. Such high-resolution structural information is available only for the fraction of the druggable proteome. At 90% sequence identity, Figure 1 shows that the coverage of the human kinome by protein crystal structures from the PDB ³⁴ is ~20%. On the other hand, the popularity of kinase inhibitors as novel therapeutics has significantly increased. Since 1995, when one of five published papers on inhibitor development was related to kinases, the interest in kinase inhibitors has grown significantly; in 2008, approximately one-third of publications reporting on inhibitor development can be linked to protein kinases (Figure 1, inset). This evident trend in pharmaceutical research creates a great demand for the structural data that would cover the entire human kinome. The gap between the availability of protein sequences and structures can be filled by protein structure prediction, particularly comparative modeling ^{35, 36}. For a target sequence, given a set of evolutionarily related protein structures, state-of-the-art template-based algorithms can construct a model whose quality is often comparable to that of a low-resolution experimentally determined structure ³⁷. However, despite having the correct global topology, theoretically predicted protein structures may still have significant structural inaccuracies in their ligand binding regions. It has been demonstrated that even moderate structural errors in the backbone and side chain coordinates interfere with traditional ligand docking approaches and cause a critical deterioration in the ability to accurately reproduce binding poses ^{32, 33}.

Figure 1.

Availability of the ligand-bound and ligand-free crystal structures for the human kinome. Inset: Histogram of the number of abstracts published since 1995 selected from the PubMed using following queries: (“inhibitor”[Text Word]) AND (“YEAR/01/01”[Publication Date]: “YEAR/12/31”[Publication Date]) and ((“inhibitor”[Text Word]) AND (“kinase”[Text Word])) AND (“YEAR/01/01”[Publication Date]: “YEAR/12/31”[Publication Date]).

On that account, the use of protein models as target receptors for ligand docking in structure-based drug development requires appropriate computational techniques that may be different from those designed to operate on the crystal structures. The recently developed FINDSITE/Q-dock ligand homology modeling (LHM) methodology is one such approach that has been demonstrated to exhibit the desired tolerance to receptor structure deformation ^{38, 39}. Conceptually similar to protein comparative modeling, LHM extends template-based techniques to the modeling of protein-ligand interactions and provides a detailed functional annotation of the target proteins. As schematically depicted in Figure 2, following protein structural characterization, the functional characterization can be considered as a three-stage process. First, functional relationships between proteins are detected by sensitive methods such as sequence profile-driven threading ^{40, 41} in order to identify essential features associated with ligand binding, i.e. functionally important residues, common molecular substructures in binding ligands and the structural conservation of their binding modes ³⁹. These insights are subsequently exploited during the initial docking of ligands by a similarity-based approach ^{39, 42}. Finally, drug candidates placed into the target binding pockets are subject to a refinement procedure to optimize the interactions with the protein and to rank the predicted poses ^{38, 43}. To deal with the problem of structural deformations when protein models are used as the target structures, low-resolution ranking and scoring techniques have been developed ^44–46.

Figure 2.

Hierarchical approach to structural and functional characterization of proteins using homology modeling techniques.

In this study, we present the results of the large-scale structure modeling and virtual screening of the entire human kinome. All-atom structural models of all kinase domains in humans have been constructed by a state-of-the-art protein structure prediction approach ^{40, 41, 47, 48}. Next, ATP-binding pockets were identified and used as the target sites in ligand-based virtual screening against a large (>2×10⁶) collection of commercially available drug-like compounds ⁴⁹ followed by ligand docking/refinement applied to the top 1×10⁴ molecules for each kinase. Ligand homology modeling ^{38, 39} produced >1×10⁹ molecular fingerprint-based similarity assessments of drug-kinase pairs and >5×10⁶ 3D models of drug-kinase complexes. The latter were subsequently evaluated by various scoring functions and finally, the ranked lists of compounds were compiled for each human kinase. Modeling accuracy is validated for protein structure prediction, binding residues identification and ligand docking using available experimental data. Compound ranking is assessed in retrospective benchmarks against several commonly used ligand libraries, including BindingDB ⁵⁰, MDL Drug Data Report ⁵¹ and the Directory of Useful Decoys ⁵². Furthermore, in a case study, we discuss the possible application of machine learning on virtual screening data to support the development of isoform-specific protein kinase inhibitors.

The full set of modeled protein structures, docked ligand conformations and compound rankings are freely available to the academic community via a user-friendly web interface that can be accessed from http://cssb.biology.gatech.edu/kinomelhm/as well as from the ZINC website (http://zinc.docking.org/applications/2010Apr/Brylinski-2010.tar.gz).

2. MATERIALS AND METHODS

2. 1. Kinase structure modeling

The sequences of all kinase domains identified in the human genome were taken from ¹. This repository contains 516 putative protein kinase genes; 409 of which are grouped into 8 major kinase families (AGC, CAMK, CK1, CMGC, RGC, STE, TK and TKL), 82 are classified as “others” and 25 are considered atypical. Protein structure modeling was carried out as follows: First, for each kinase domain structure templates were selected from a non-redundant template library by our threading algorithm PROSPECTOR_3 ^{40, 41}, which was designed to detect close as well as remote homologous templates. Subsequently, threading templates were submitted to TASSER ^{47, 48}, a coarse-grained structure assembly/refinement procedure guided by tertiary restraints extracted from the template structures. All-atom models were constructed from Cα coordinates obtained from the TASSER simulations by PULCHRA ⁵³. Finally, the kinase structures were energy minimized in the CHARMM22 force field ⁵⁴ using the Jackal modeling package ⁵⁵. Modeled kinase structures were then taken as targets for the prediction of ATP-binding sites by FINDSITE ^{56, 57}, a threading-based method that identifies ligand-binding sites based on binding site similarity among superimposed groups of functionally and structurally related template structures. The ATP-binding pockets were used as the target sites to dock ligands.

2.2. Ligand docking and ranking

The ligand docking procedure consisted of initial ligand placement by FINDSITE^{LHM 39} followed by low-resolution refinement by Q-Dock^{LHM 38} and all-atom refinement using AMMOS ⁵⁸. FINDSITE^LHM is a fast ligand homology modeling approach that docks flexible ligands by a simple superpositioning procedure. It uses a collection of template-bound ligands extracted from binding sites predicted by FINDSITE to derive the common molecule substructures, viz. the anchor functional groups. Subsequently, the consensus binding poses of the anchor substructures are used for target ligand superposition, where the flexibility of a ligand is accounted for by the superposition of multiple low-energy conformations generated by BALLOON ⁵⁹. The conformation that can be superimposed onto the reference coordinates with the lowest RMSD structure to the predicted anchor pose is selected as the final model. Initial binding poses generated by FINDSITE^LHM were submitted to low-resolution refinement by Q-Dock^LHM. Q-Dock^LHM is a direct extension of Q-Dock ⁴⁴ that additionally includes harmonic RMSD restraints imposed on the predicted anchor-binding pose. The lowest-energy conformation generated during the Replica Exchange Monte Carlo sampling was selected as the final docking result. Ligand poses provided by Q-Dock^LHM were transformed into the all-atom representation and further refined by molecular mechanics optimization using AMMOS ⁵⁸. AMMOS uses the AMMP molecular simulation package ⁶⁰ to carry out automatic refinement of the protein-ligand complexes. We used the sp4 force field in all simulations; protein atoms within a 12 Å sphere around the ligand were allowed to be flexible (AMMOS Case 4).

To provide compound ranking in virtual screening, we applied the following scoring functions: ligand-based molecular fingerprints implemented in FINDSITE ^{56, 61}, anchor substructure coverage, where the anchor substructures were identified by FINDSITE^{LHM 39}, structure-based scoring by the total energy and the pocket-specific component from Q-Dock^LHM’s force field ³⁸ and the total docked energy provided by AMMOS ⁵⁸.

2.3. Datasets

2.3.1. ZINC

Each protein kinase was screened against 2,095,759 compounds from the ZINC7 library ⁴⁹. In the first step, a fast ligand-based screening was applied using molecular fingerprints provided by FINDSITE ^{56, 57}, as described above. Subsequently, for each target, the top 10,000 compounds (0.5% of the library) were selected based on the modified Tanimoto score ^{39, 62, 63} and submitted to molecular docking by FINDSITE^LHM followed by Q-Dock^LHM and AMMOS. Finally, the compounds were re-ranked by the structure-based scoring functions.

2.3.2. PDB

Protein structure modeling, binding residue prediction and docking accuracy were assessed for 326 kinase crystal structures taken from ⁶⁴. The dataset consists of 57 different human kinases with a ligand bound in the ATP-binding site (278 unique protein-ligand pairs) and 48 ligand-free forms.

Kinase structure modeling accuracy was assessed by the global Cα RMSD and the TM-score ⁶⁵. Local structural distortions of the binding pockets were evaluated by their Cα and all-atom RMSD calculated over the binding residues identified by LPC ⁶⁶. The accuracy of ATP-binding site detection by FINDSITE was expressed as the distance of the predicted site from the ligand geometric center in the crystal structures and the Matthew’s correlation coefficient (MCC) calculated for the binding residues:

$$\mathit{MCC}=\frac{TP\times TN-FP\times FN}{\sqrt{(TP+FP)\times (TP+FN)\times (TN+FP)\times (TN+FN)}}$$ (Eq. 1)

where TP, TN, FP and FN denote respectively: true positives (correctly predicted binding residues), true negatives (residues correctly predicted not to bind a ligand), false positives (overpredicted binding residues) and false negatives (missing binding residues).

To evaluate docking accuracy, we use the fraction of correctly predicted binding residues as well as the fraction of recovered native specific protein-ligand contacts ³⁸. In theoretical protein models, the local geometry of the binding pocket frequently deviates from the experimental structure. Therefore, ligand poses transferred from the crystal structures upon the superposition of the binding residues roughly estimate the upper bound for ligand docking accuracy against protein models. Ligands randomly placed into the ATP-binding pockets within a distance of 7 Å (docking sphere) from the predicted pocket center delineate the lower bound of docking accuracy.

2.3.3. BindingDB

Ranking accuracy in virtual screening was assessed for 362 known active compounds selected from BindingDB ⁵⁰. The top 10,000 compounds from virtual screening against the ZINC7 library were used as background compounds. For each known kinase inhibitor, we assess the improvement of ranking by structure-based scoring using Q-Dock^LHM and AMMOS over the fingerprint-based scoring by FINDSITE.

2.3.4. KEGG

The rank of ATP for each kinase target was calculated versus 12,158 background molecules from the KEGG compound library ⁶⁷.

2.3.5. DUD

The Directory of Useful Decoys ⁵² was designed for benchmarking virtual screening approaches and contains 40 protein targets, 2,950 active compounds and 36 decoy molecules per one active compound with similar physical properties. Seven targets from DUD belong to the human kinase family: CDK2, EGFR, FGFR1, KDR, p38a, PDGFRb and SRC. Here, we use these targets to provide a comparative assessment of the screening protocols used in this study and in state-of-the-art virtual screening using DOCK ⁶⁸. The energy-based ligand rankings by DOCK3.5 applied to the crystal structures of the target kinases were taken from ⁵². In addition, we carried out docking simulations using DOCK6 against the crystal as well as modeled kinase structures. Target receptor structures were prepared by Chimera ⁶⁹ using the default set of parameters. Ligand preparation including the Gasteiger-Marsili partial charge assignment and the calculation of hydrogen positions were done using OpenBabel ⁷⁰. Binding poses generated by flexible ligand docking simulations using a default “anchor and grow” protocol were ranked by the total grid score. The results provided by DOCK3.5/6 were compared to ligand rankings obtained by low-resolution docking/scoring by Q-Dock^{LHM 38, 44} (knowledge-based potential) and FINDSITE^{LHM 39} (anchor coverage) using modeled structures. Furthermore, we applied data fusion to combine the results from virtual screening using the pocket-specific potential (Q-Dock^LHM) and the anchor coverage (FINDSITE^LHM). Here, we use the SUM rule that is expected to be less sensitive to noisy input than both extreme rules ⁷¹ and is preferred when fusion is by rank ⁷². For a given library compound k, a combined score (CS) is calculated from:

$${CS}^k=\sum _{i=1}^n{r}_i$$ (Eq. 2)

where n is the number of ranked lists (in our case, n=2: Q-Dock^LHM and FINDSITE^LHM) and r_i denotes the rank position of the library compound k in the i-th ranked list.

The performance of DOCK3.5/6 and Q-Dock^LHM/FINDSITE^LHM in virtual screening for kinase inhibitors is assessed by EF₁₀ (enrichment factor calculated for the top 10% of the ranked screening library) ^{39, 73}, BEDROC20 (Boltzmann-enhanced discrimination of ROC) ⁷³, AUAC (area under the accumulation curve) ⁷³ and ACT-50% (the top fraction of ranked library that contains 50% of the active compounds). Random ligand ranking yields EF₁₀, BEDROC20, AUAC and ACT-50% of 1.0, 0.1, 0.5 and 0.5, respectively.

2.3.6. MDDR

MDL Drug Data Report provides comprehensive information on bioactive compounds compiled from published and unpublished sources ⁵¹. 562 protein kinase C (PKC) inhibitors were selected from MDDR (MDL activity index: 78374) and used in virtual screening against 9 isoenzymes of PKC: α, β, γ, delta;, ε, η, θ, ι and ζ. For each PKC isoform, 10,000 compounds randomly selected from the ZINC7 database ⁴⁹ were used as the background library.

2.3.7. PKC

In addition to the assessment of the ligand ranking capability for protein kinase C, we also investigated the possibility of the prediction of inhibitor specificity toward different isoenzymes of PKC by a machine learning approach. Here, we use 10 inhibitors collected from the literature, for which half-maximal inhibition constants (IC₅₀) values toward PKC isoforms were determined experimentally: corallidictyal ⁷⁴, GF-109203X ⁷⁵, Gö-6976 ⁷⁶, JTT-010 ⁷⁷, K252a ⁷⁸, midostaurin ⁷⁹, rottlerin ⁸⁰, ruboxistaurin ⁸¹, staurosporine ⁸² and UCN-01 ⁸³. A simple three-state classification model was constructed; for each PKC isoenzyme, the inhibitors were divided into three classes based on the IC₅₀ values: class I, good binders (IC₅₀ < 100 nM), class II, weak binders (100nM < IC₅₀ < 1 μM) and class III, non-binders (IC₅₀ > 1 μM). The Supporting Vector Machine (SVM, nu-SVC type with a polynomial kernel) ⁸⁴ was trained on the following features: docking scores (raw score and the Z-score from virtual screening): fingerprint-based (FINDSITE), final docked energy (Q-Dock^LHM), pocket specific component (Q-Dock^LHM), and the chemo-physical properties of the inhibitors: molecular weight (MW), octanol/water partition coefficient (logP) and topological polar surface area (PSA). The molecular properties were calculated by OpenBabel ⁷⁰. The classification model was validated using the following leave-one-out procedure: in each round, one inhibitor was removed from the dataset, the SVM model was trained on the inhibition data for the remaining compounds and the excluded inhibitor was assigned a binding class for each PKC isoenzyme. The accuracy is assessed in terms of the fraction of correct assignments. Finally, the SVM model was trained on all experimental data and the prediction was made for PKC isoenzyme–inhibitor pairs for which no inhibition constants are reported in the literature.

3. RESULTS

3. 1. Modeled structures for the human kinome

Template-based modeling is one of the most frequently used techniques in protein structure prediction and has the capability of providing reliable models in the presence of evolutionarily related template structures ^{35, 36}. In this study, we constructed structure models for all kinase sequences identified in the human kinome by our protein structure prediction protocol: threading by PROSPECTOR_3 ^{40, 41} followed by structure assembly/refinement using TASSER ^{47, 48}. Figure 3 presents the global Cα root-mean-square-deviation, RMSD, TM-score ⁶⁵ and binding pocket RMSD from the crystal structure for the set of 57 ligand-bound and 48 ligand-free human kinases ⁶⁴ that have experimentally determined structures in the PDB. The global structures of kinase domains have an average Cα RMSD (TM-score) from the holo and apo crystal structures of 2.75Å (0.92) and 3.13Å (0.90), respectively. The lower RMSD and higher TM-score values calculated for holo vs. apo structures reflect the fact that most of the template structures in the PDB are in the ligand-bound functional state (see Figure 1) and the force field used by TASSER for structure refinement favors conformations that are typically more compact and contain more inter residue contacts than the open conformational states. Figure 3C shows the local deviations from the experimental structure for ATP-binding pockets; the accuracy of these regions is critical for ligand docking and ranking. The average Cα (all-atom) RMSD calculated over the binding residues is 1.27 Å (2.36 Å). Despite progress in the prediction of residue rotamers ^85–87, side chain modeling still needs further improvement. Nevertheless, these values concur with the estimated plasticity of the binding sites that have the capability to bind the same ligand (or class of ligands) in the kinase family ⁸⁸ and proteins in general ³⁹. In contrast to many ligand-docking algorithms that require highly accurate experimental structures, the local distortions of ligand binding regions are tolerated to some extent by docking approaches that use a lower resolution description ^{38, 44–46}.

Figure 3.

Accuracy of kinase structure modeling using TASSER. Global Cα RMSD (A) and TM-score (B) are calculated versus ligand-bound (holo) and ligand-free (apo) structural forms of the target proteins. Local Ca and all-atom RMSD calculated over the binding residues are shown in C.

3. 2. ATP-binding pocket prediction by FINDSITE

To dock ligands into the modeled kinase structures, we used binding pockets predicted by FINDSITE, a threading-based binding site prediction/protein functional inference/ligand screening algorithm that detects common ligand binding sites in a set of evolutionarily related proteins ^{56, 57}. The average number of binding sites predicted by FINDSITE for a kinase target is 32. Here, we use only the top-ranked pockets with the majority of low ranked sites likely involved in nonspecific ligand binding. The results of ATP-binding pocket prediction carried out for 57 different human kinases and 278 ligands are shown in Figure 4. Considering a cutoff distance of 4 Å as the hit criterion, the success rate for all complexes and for a non-redundant set with respect to the protein sequences is 86.7% and 94.7%, respectively. In most of the cases, the predicted distance is less than 2.5 Å. This very high accuracy of binding site prediction results in high Matthew’s correlation coefficients (MCC) calculated for the binding residues; for most of the complexes, the MCC is >0.80 (Figure 4, inset). Two major factors account for the exceptional efficiency of ATP-binding site detection: the kinase structures have been modeled by TASSER to very high accuracy and most of the currently available kinase inhibitors, whose complexes are present in the PDB ³⁴, target ATP-binding sites ^{64, 89}.

Figure 4.

ATP-binding pocket detection by FINDSITE. The results are presented as the cumulative fraction of kinase targets with a distance between the center of mass of an inhibitor in the crystal complex and the center of the predicted binding sites, less than or equal to the distance displayed on the x axis. Open circles show the results for a non-redundant (nr) dataset with respect to the target proteins. Gray area corresponds to randomly selected patches on the protein surface. Inset: Matthew’s correlation coefficient calculated for the predicted binding residues.

3.3. Ligand binding pose prediction

Low-resolution docking techniques are frequently used to dock ligands into the distorted binding sites of the modeled receptor structures ^{38, 44–46}. In Figure 5, we assess the accuracy of ligand docking into the ATP-binding sites of modeled kinase structures for 278 unique protein-ligand pairs using FINDSITE^LHM, Q-Dock^LHM and an all atom refinement procedure, AMMOS ⁵⁸. The upper bound for docking accuracy is estimated by transferring ligands from the crystal structures into the modeled structures upon the local superposition of the binding residues. The fraction of correctly predicted binding residues (Figure 5A) is the highest for Q-Dock^LHM and is very close to the estimated upper bound. All-atom refinement by AMMOS recovers less binding residues, and is comparable in performance to FINDSITE^LHM. The fraction of correctly predicted specific protein-ligand contacts, (essential for effective ligand ranking), provides a more detailed assessment of the docking accuracy. Previous benchmark simulations demonstrated that ligand homology modeling by FINDSITE^LHM followed by an anchor-constrained low-resolution refinement by Q-Dock^LHM outperforms other approaches in ligand binding pose prediction against modeled receptor structures ³⁸. Figure 5B shows that FINDSITE^LHM provides an approximately correct binding pose, which is subsequently improved by low-resolution refinement using Q-Dock^LHM. This procedure recovers significantly more specific protein-ligand contacts than all-atom refinement using AMMOS. It is noteworthy that all programs used for ligand docking perform significantly better than random ligand placement in terms of the recovered binding residues as well as the specific protein-ligand contacts.

Figure 5.

Docking accuracy of the ligand homology modeling approach applied to the human kinome. Fraction of binding residues (A) and specific protein-ligand contacts (B) predicted by FINDSITE^LHM, Q-Dock^LHM and AMMOS is compared to the ligand poses directly transferred from the crystal structures as well as to ligands randomly placed into the binding pockets.

The success of a refinement procedure depends on the quality of the force field used. The latter can be assessed by the correlation between the native-likeness, e.g. RMSD from the crystal ligand binding pose and the energy score, and the location of the energy minimum; the lowest energy pose should correspond to a conformation close to native. Here, for four representative examples, we evaluate the quality of the Q-Dock^LHM’s force field that impacts refinement outcome. In Figures 6A for cyclin-dependent kinase 2, CDK2, and in Figure 6B for proto-oncogene serine/threonine protein kinase, PIM1, we show that when the docking energy score is well correlated with RMSD and the energy minimum is located close to the ligand-binding pose in the crystal structure, not surprisingly, low-resolution refinement improves docking results; the fraction of specific contacts increases from 0.65 (using FINDSITE^LHM) to 0.70 (using Q-Dock^LHM) and from 0.45 to 0.60 respectively. On the other hand, in some cases, the energy score is not correlated with the native-likeness of the ligand poses; this results in minor (from 0.41 to 0.50 of the fraction of specific native contacts that are recovered for tyrosine kinase FGFR2, Figure 6C) or no improvement by Q-Dock^LHM over FINDSITE^LHM (0.40 for both methods for CDK2, Figure 6D). Nevertheless, significantly better ligand binding poses are generated by Q-Dock^LHM for most of the modeled complexes, which is critical for ligand ranking. As shown in Figure 5B, the fraction of complexes with 0.40, 0.50, 0.60 and 0.70 of the specific native contacts recovered by low-resolution, Q-Dock^LHM, refinement is 0.83, 0.72, 0.56 and 0.30, respectively.

Figure 6.

Low-resolution docking/refinement by ligand homology modeling using protein models as the target receptors. A – CDK2, 1oiq; B – PIM1, 1yxx; C – FGFR2, 1oec and D – CDK2, 2btr. Left, middle: Inhibitor binding poses predicted by FINDSITE^LHM and Q-Dock^LHM (solid sticks, colored by atom type) are compared to the crystal structures (transparent sticks). Protein models (binding residues colored in red) are superposed onto the crystal structures of the target kinases (binding residues colored in orange). Right: correlation of the Q-Dock energy score and RMSD from the crystal binding pose for the ligand conformations sampled using Replica Exchange Monte Carlo (REMC). The red line highlights low-energy conformations for the broad range of RMSD values.

We next consider some specific examples:

3.4. Staurosporine binding mode in modeled kinase structures

A natural product of S. staurosporeus, staurosporine (STU), was first described as an inhibitor of protein kinase C ⁸². Later on, STU was demonstrated to have nanomolar potency toward a variety of other protein kinases ^{90, 91}. STU non-selectively inhibits protein kinases by competitively binding to the ATP-binding site. Highly conserved across the protein kinase family, the position of STU in the ATP-binding pocket (see Figure 7) is stabilized by predominantly hydrophobic interactions and hydrogen bonds ^{92, 93}. The inhibitor mimics several aspects of adenosine binding; the lactam ring of STU occupies a similar position to the amino group of ATP and the sugar moiety of STU binds to the region occupied by the ribose of ATP, pointing out of the binding site. Despite the structural distortions of ATP-binding sites in modeled kinase structures (see Figure 3C), similar binding modes of STU and ATP were recovered by the low-resolution docking using Q-Dock^LHM. This is shown in Figure 8 for nine protein kinases whose crystal structures are not available in the PDB ³⁴. High accuracy of STU docking into the ATP-binding sites of homology models has been reported previously for eight protein kinases ⁸⁸. Furthermore, it is noteworthy that structure-based virtual screening against protein models using the pocket-specific potential as a scoring function assigned very high Z-scores and corresponding ranks to both compounds (Figure 8). This high ranking efficiency is encouraging since staurosporine, as a potent and promiscuous kinase inhibitor, represents a prototypical ATP-competitive lead compound ⁹⁴.

Figure 7.

Crystal structures of several protein kinases complexed with staurosporine (STU) and ATP. A – CDK2 (STU: 1aq1, ATP: 1b38), B – GSK3B (STU: 1q3d, ADP: 1j1c), C – LCK (STU: 1qpd, ANP: 1qpc), D – PIM1 (STU: 1yhs, AMP: 1yxu), E – PDK1 (STU; 1oky, ATP: 1h1w), F – MAPKAPK2 (STU: 1nxk, ADP: 1ny3). STU, the set ATP/ADP/AMP/ANP and selected binding residues are colored in green, red and blue, respectively.

Figure 8.

Modeled structures of protein kinases bound to staurosporine (STU) and ATP. A – CDC2, B – Erk1, C – FGR, D – LYN, E – PKACa, F – PKCa, G – PKCg, H – PKG1, I – smMLCK. STU, ATP and selected binding residues are colored in green, red and blue, respectively. ATP and STU ranks and Z-scores from virtual screening using Q-Dock^LHM against modeled kinase structures are given.

3.5. Ligand ranking

The goal of virtual screening is to rapidly assess a large library of compounds in order to identify those molecules that most likely bind to a drug target. To estimate the reliability of ligand ranking, known active molecules are typically included in the screening library; high ranks assigned to these compounds by a virtual screening approach indicate that the top fraction of the ranked library is significantly enriched in biologically active compounds. Here, we assess the accuracy of ligand- and structure-based virtual screening for a set of 362 known kinase inhibitors selected from the BindingDB ⁵⁰. We note that only compounds that are not present in the PDB ³⁴ are used in this analysis. The results in terms of the ranks assigned to known active molecules in the screening library of the top 10,000 ranked compounds of the ZINC7 library are presented in Figure 9. First, we assess the improvement in ligand ranking of structure-based over ligand-based virtual screening. For most of the compounds, docking-based scores provide better (lower) ranks than the fingerprint-based scoring using FINDSITE, with the low-resolution scoring by Q-Dock^LHM providing the most effective ligand ranking. The number of compounds assigned with ranks <100 (the top 1% of the library) is 3, 68 and 2 for FINDSITE, Q-Dock^LHM and AMMOS, respectively. Q-Dock^LHM assigned ranks lower than 1,000 (the top 10% of the library) to almost twice as many known inhibitors as AMMOS and four times more inhibitors than FINDSITE. Separately, we assess the ranking of ATP that binds to all kinases (Figure 9, inset). For 95% of the protein kinases, ATP was ranked by Q-Dock^LHM within the top 1% of the screening library. Strong evolutionary relationships between protein kinases are easily detected by sequence profile-driven threading; this results in similar sets of templates identified for individual members. Hence, the ranks assigned to ATP by FINDSITE using the molecular fingerprints extracted from template-bound ligands are invariant across the kinase family. The improved ranking provided by Q-Dock^LHM over FINDSITE provides a very strong justification for the more CPU-expensive Q-Dock^LHM-based ligand docking. We note that the top 10,000 compounds selected by FINDSITE from the ZINC7 database ⁴⁹ have been re-ranked by Q-Dock^LHM and AMMOS for all 516 kinases identified in the human proteome.

Figure 9.

Performance of virtual screening on the BindingDB dataset. Active compounds are sorted by increasing rank reported by FINDSITE fingerprints (ligand-based screening), Q-Dock^LHM (structure-based screening, low-resolution) and AMMOS (structure-based screening, high-resolution). Inset: ATP ranks for all protein kinases; for FINDSITE, the ranks in the KEGG compound library are used.

3. 6. Performance on the DUD dataset

The Directory of Useful Decoys (DUD) provides a large unbiased benchmark set to test the performance of virtual screening approaches ⁵². In contrast to many other datasets, the decoy compounds included in DUD are physically similar to active compounds, yet they have a different topology from their active counterparts. This important feature helps avoid the artificial enrichment often seen in virtual screening studies ⁹⁵; hence DUD is frequently used in the assessment of the performance of virtual screening approaches ^96–100. In Table 1, we compare the performance of the ligand homology modeling approach (FINDSITE^LHM/Q-Dock^LHM) used in this study to DOCK3.5/6, the all-atom docking/screening tool on a set of 7 protein kinases from DUD. First, we note that for receptor crystal structures, DOCK6 provides higher enrichment with respect to the previous version, DOCK3.5. In benchmarks against modeled structures, considering single scoring functions, FINDSITE^LHM performs better on average than DOCK6, Q-Dock^LHM and AMMOS with an average EF₁₀, BEDROC20, AUAC and ACT-50% (the top fraction of ranked library that contains 50% of the active compounds) of 1.905, 0.133, 0.625 and 0.285, respectively. Moreover, the performance of FINDSITE^LHM for protein models is close to or depending on the metric used exceeds the performance of DOCK6 applied to the crystal structures, 1.955, 0.173, 0.383 and 0.779. The two docking algorithms, DOCK6 and Q-Dock^LHM perform quite comparably against modeled structures; DOCK6 outperforms Q-Dock^LHM with respect to EF₁₀ and BEDROC20; however, the average AUAC and ACT-50% are notably better for Q-Dock^LHM. Poor AUAC and ACT-50% measures calculated for ligands ranked by DOCK6 suggest that active compounds are not equally well distributed across the screening library and low ranks are assigned to a significant fraction of known inhibitors. In addition, we find that high-resolution refinement and scoring using AMMOS applied to ligand poses generated by Q-Dock^LHM does not improve ligand ranking. The combined approach, data fusion using the SUM rule applied to ligand rankings from FINDSITE^LHM and Q-Dock^LHM, performs significantly better than the other approaches used in this study and yields an average EF₁₀, BEDROC20, AUAC and ACT-50% of 2.378, 0.162, 0.624 and 0.316, respectively. The most important conclusion emerging from this study is that ligand homology modeling by FINDSITE^LHM/Q-Dock^LHM using predicted protein structures is a competitive alternative to classical structure-based virtual screening with better or at least comparable efficacy in ligand ranking to approaches that require solved protein crystal structures with bound ligands.

Table 1.

Performance of ligand homology modeling on seven protein kinases from the DUD dataset compared to the results obtained using DOCK. Ranking capability is assessed by the enrichment factor (EF₁₀), Boltzmann-enhanced discrimination of ROC (BEDROC20), the area under the accumulation curve (AUAC) and the top fraction of ranked library that contains 50% of the active compounds (ACT-50%).

		CDK2	EGFR	FGFR1	KDR	p38a	PDGFRb	SRC	Average ±SD
DOCK3.5 crystal structures	BEDROC20	0.189	0.200	0.003	0.085	0.115	0.009	0.026	0.090 ±0.082
	EF10	2.200	2.545	0.085	1.081	1.992	0.197	0.323	1.203 ±1.038
	AUAC	0.549	0.565	0.201	0.402	0.532	0.323	0.448	0.431 ±0.134
	ACT-50%	0.340	0.274	0.885	0.682	0.463	0.742	0.494	0.554 ±0.223
DOCK6 crystal structures	BEDROC20	0.250	0.236	0.107	0.198	0.106	0.161	0.150	0.173 ±0.058
	EF10	2.600	2.590	1.453	1.892	1.289	1.987	1.871	1.955 ±0.504
	AUAC	0.459	0.441	0.346	0.393	0.314	0.355	0.371	0.383 ±0.052
	ACT-50%	0.586	0.717	0.847	0.760	0.868	0.861	0.812	0.779 ±0.101
DOCK6 protein models	BEDROC20	0.104	0.255	0.003	0.119	0.070	0.306	0.090	0.135 ±0.107
	EF10	1.000	2.568	0.085	1.351	0.898	2.930	1.161	1.428 ±0.992
	AUAC	0.341	0.433	0.196	0.399	0.236	0.404	0.290	0.328 ±0.091
	ACT-50%	0.781	0.725	0.863	0.711	0.923	0.832	0.911	0.821 ±0.085
AMMOS protein models	BEDROC20	0.049	0.014	0.058	0.038	0.069	0.072	0.033	0.048 ±0.021
	EF10	1.400	0.158	0.932	0.405	1.055	1.338	0.581	0.838 ±0.472
	AUAC	0.671	0.466	0.611	0.422	0.510	0.475	0.518	0.525 ±0.087
	ACT-50%	0.299	0.537	0.350	0.590	0.501	0.527	0.427	0.462 ±0.107
Q-DOCKLHM protein models	BEDROC20	0.163	0.062	0.105	0.076	0.088	0.099	0.020	0.088 ±0.044
	EF10	2.400	0.968	1.610	1.081	1.289	1.210	0.194	1.250 ±0.668
	AUAC	0.665	0.553	0.613	0.533	0.526	0.577	0.466	0.562 ±0.064
	ACT-50%	0.225	0.418	0.364	0.423	0.477	0.438	0.529	0.411 ±0.097
FINDSITELHM protein models	BEDROC20	0.155	0.067	0.175	0.146	0.125	0.113	0.151	0.133 ±0.035
	EF10	2.000	1.014	2.712	2.027	1.797	1.656	2.129	1.905 ±0.515
	AUAC	0.690	0.525	0.686	0.563	0.595	0.618	0.700	0.625 ±0.069
	ACT-50%	0.204	0.442	0.186	0.372	0.249	0.317	0.228	0.285 ±0.095
Data fusion protein models	BEDROC20	0.321	0.107	0.210	0.114	0.127	0.161	0.096	0.162 ±0.080
	EF10	4.400	1.689	2.966	1.486	1.875	2.229	2.000	2.378 ±1.010
	AUAC	0.724	0.552	0.698	0.565	0.584	0.627	0.619	0.624 ±0.066
	ACT-50%	0.155	0.420	0.184	0.420	0.391	0.314	0.328	0.316 ±0.109

3. 7. Virtual screening for isoform-specific PKC inhibitors

An early event in signal transduction pathways, the activation of the protein kinase C family (PKC), leads to many biological responses that regulate major cellular functions ¹⁰¹. Different PKC isoenzymes are considered to be promising targets in the treatment of many diseases, including diabetes, multiple sclerosis, cardiovascular disease, cancer and Alzheimer’s ^{5, 6, 8}. Based on their structure and regulation mechanisms, the isoforms of protein kinase C can be divided into three categories: conventional calcium-dependent PKCs (α, β_I, β_II and γ) that are activated by both phospholipids and diacylglycerol (DAG), novel PKCs (δ, ε, η and θ) that require phospholipids and DAG for activation but do not require Ca²⁺ and atypical PKCs (ι/λ and ζ) that are unresponsive to both activators ^{102, 103}. Most of the compounds inhibit PKC isoforms non-selectively; to exploit the distinct function of different PKC isoenzymes, isoenzyme-specific inhibitors are highly desired. Here, in a benchmark scenario, we demonstrate how virtual screening data can be used to support the development of isoform-specific PKC inhibitors.

In the first step, we carried out the retrospective evaluation of the virtual screening for the PKC inhibitors using 562 active compounds from the MDDR database ⁵¹ and 10,000 random decoys from the ZINC7 library ⁴⁹. We note that MDDR does not specify the selectivity of PKC inhibitors toward different isoenzymes. Therefore, the results in terms of the enrichment behavior plots are presented in Figure 10 for each isoform of the PKC. This example shows that the compound ranking using an all-atom scoring function such as the one used by AMMOS ⁵⁸ is ineffective when modeled protein structures are used as the target receptors. It has been already demonstrated in more representative benchmarks that all-atom approaches for ligand docking and ranking are highly sensitive to structural distortions in ligand binding regions ^{38, 39, 44}. Molecular fingerprints provided by FINDSITE perform better that random ligand selection with 4.8% and 24.0% of the known inhibitors recovered in the top 1% and 10% of the screening library, respectively. Since PKC isoforms are closely related to each other, the ranks of library compounds by FINDSITE are identical for all isoenzymes; similar behavior was seen when FINDSITE is applied to the prediction of ATP binding (see Figure 9, inset), as FINDSITE emphasizes the conserved binding features across a protein family; here, we are interested in their differences. Quite similar performance is observed for structure-based virtual screening by the total energy reported by Q-Dock^LHM (which includes both generic and protein specific components, see Methods, below) Here, the percentage of active compounds recovered in the top 1% (10%) of the library varies from 2.8% (12.6%) for PKC-γ to 10.1% (27.6%) for PKC-ι. Undoubtedly, the best performance is obtained using the pocket-specific component of the Q-Dock^LHM’s force field as a scoring function to rank ligands. The fraction of known PKC inhibitors ranked within the top 1% and 10% of the library varies from 11.7% (PKC-α) to 13.9% (PKC-ι) and from 34.9% (PKC-α) to 42.3% (PKC-ε), respectively. Furthermore, using the pocket specific scoring function, ligand ranking is very stable across different isoforms of the PKC.

Figure 10.

Virtual screening for protein kinase C inhibitors. The enrichment behavior for FINDSITE (molecular fingerprints), Q-Dock^LHM (total energy score and the pocket-specific component) and AMMOS (all-atom scoring) is compared to a random ligand selection for different isoenzymes of PKC.

Next, we employed a simple machine learning model to demonstrate that virtual screening data can be used for the prediction of the inhibitor specificity toward different PKC isoenzymes. Leave-one-out cross validation (Table 2, in italics) shows that for 7 out of 10 inhibitors (GF-109203X, Gö-6976, K252a, midostaurin, rottlerin, staurosporine and UCN-01) the three-state binding assignment of good, weak and non-binders (see Materials and Methods) was better than random (random accuracy is 33.3%). The highest benchmark accuracy (60%) is observed for the indolocarbazole Gö-6976, which is the first discovered PKC inhibitor that was shown in vivo to discriminate between Ca²⁺-dependent and Ca²⁺-independent PKC isoenzymes ⁷⁶. In the validation of our model, Gö-6976 is predicted to inhibit α and β isoforms with high affinity <100 nM (experimental IC₅₀ values are 2.3 nM and 6.2 nM, respectively). PKC isoenzymes d and e are false positives i.e. predicted to be inhibited, while the experimental data shows no inhibition. Gö-6976 is correctly assigned as a non-active compound against the isoform ζ. The activity of three other Ca²⁺-independent PKC isoenzymes, η, θ and ι, is also predicted to be unaffected by Gö-6976; this is in good agreement with its class-selective inhibition profile. Another interesting example is rottlerin that was predicted as a weak/non-inhibitor for most PKC isoforms. In the recent study of protein kinases and inhibitors, rottlerin failed to show any PKC inhibitory activity against the α and delta; PKC isotypes ^{104, 105}, which is consistent with our results. Considering the relatively high prediction accuracy, we used all experimental data to predict IC₅₀ values for PKC isoenzyme–inhibitor pairs for which no inhibition constants are reported in the literature (Table 3, in bold).

Table 2.

Benchmarking results for the prediction of the inhibitor selectivity toward protein kinase C isoenzymes. Experimental and benchmark values of IC₅₀ are shown in regular font and italics, respectively. Correct and incorrect classifications are highlighted in green and red, respectively.

Inhibitor	IC50 values for PKC isoenzymes
	α (50%)	β (75%)	γ (71%)	δ (11%)	ε (30%)	η (80%)	θ	ι	ζ (20%)
corallidictyal (0%)	30 μM 100nM – 1 μM				89 μM 100nM – 1 μM	>300 μM 100nM – 1 μM			>300 μM 100nM – 1 μM
GF-109203X (40%)	8.4 nM <100 nM	18 nM <100 nM		210 nM <100 nM	132 nM <100 nM				5.8 μM <100 nM
Gö-6976 (60%)	2.3 nM <100 nM	6.2 nM <100 nM		No inh 100nM – 1 μM	No inh <100 nM				No inh >1 μM
JTT-010 (33%)	86 nM 100nM – 1 μM	4 nM 100nM – 1 μM	110 nM 100nM – 1 μM	54 nM 100nM – 1 μM	490 nM 100nM – 1 μM				1.7 μM 100nM – 1 μM
K252a (50%)	40 nM 100nM – 1 μM		400 nM 100nM – 1 μM	925 nM 100nM – 1 μM	4.5 μM 100nM – 1 μM	490 nM 100nM – 1 μM			4.2 μM 100nM – 1 μM
midostaurin (57%)	24 nM <100 nM	17 nM <100 nM	18 nM <100 nM	360 nM <100 nM	4.5 μM <100 nM	60 nM <100 nM			>10 μM <100 nM
rottlerin (57%)	30 μM >1 μM	42 μM 100nM – 1 μM	40 μM >1 μM	6 μM 100nM – 1 μM	100 μM >1 μM	82 μM >1 μM			100 μM 100nM – 1 μM
ruboxistaurin (28%)	360 nM <100 nM	4.7 nM <100 nM	300 nM <100 nM	250 nM <100 nM	600 nM <100 nM	52 nM <100 nM			>10 μM <100 nM
staurosporine (50%)	8.7 nM 100nM – 1 μM	11 nM <100 nM	11 nM 100nM – 1 μM	4.3 nM 100nM – 1 μM	7.4 nM <100 nM				1.7 μM >1 μM
UCN-01 (50%)	29 nM <100 nM	34 nM <100 nM	30 nM <100 nM	590 nM <100 nM	530 nM <100 nM				No inh <100 nM

Table 3.

Prediction of the inhibitor selectivity toward protein kinase C isoenzymes by machine learning on virtual screening data. Experimental and predicted values of IC₅₀ are shown in regular font and bold, respectively.

Inhibitor	IC50 values for PKC isoenzymes
	α	β	γ	δ	ε	η	θ	ι	ζ
corallidictyal	30 μM	>1 μM	>1 μM	>1 μM	89 μM	>300 μM	>1 μM	>1 μM	>300 μM
GF-109203X	8.4 nM	18 nM	<100 nM	210 nM	132 nM	<100 nM	<100 nM	100nM – 1 μM	5.8 μM
Gö-6976	2.3 nM	6.2 nM	>1 μM	No inh	No inh	>1 μM	>1 μM	>1 μM	No inh
JTT-010	86 nM	4 nM	110 nM	54 nM	490 nM	100nM –1 μM	<100 nM	>1 μM	1.7 μM
K252a	40 nM	<100 nM	400 nM	925 nM	4.5 μM	490 nM	100nM – 1 μM	>1 μM	4.2 μM
midostaurin	24 nM	17 nM	18 nM	360 nM	4.5 μM	60 nM	<100 nM	>1 μM	>10 μM
rottlerin	30 μM	42 μM	40 μM	6 μM	100 μM	82 μM	>1 μM	100nM – 1 μM	100 μM
ruboxistaurin	360 nM	4.7 nM	300 nM	250 nM	600 nM	52 nM	100nM – 1 μM	100nM – 1 μM	>10 μM
staurosporine	8.7 nM	11 nM	11 nM	4.3 nM	7.4 nM	100nM – 1 μM	<100 nM	<100 nM	1.7 μM
UCN-01	29 nM	34 nM	30 nM	590 nM	530 nM	<100 nM	<100 nM	>1 μM	No inh

Finally, we apply the SVM model to assign the selectivity toward PKC isoenzymes to 562 known inhibitors from MDDR. Since no information on the selectivity profile is provided by MDDR, we indirectly validate the results using the Google search engine. The results are shown in Figure 11. Most of the compounds were predicted by the SVM to inhibit the conventional PKC isoforms with an IC₅₀ <100 nM, whereas relatively few inhibitors were predicted to be atypical PKC specific (Figures 11A and B). This trend is in good qualitative agreement with the number of hits reported by Google (Figure 11C). The highest number of hits was obtained using “protein kinase C alpha inhibitors” as the query phrase. Significantly fewer hits are reported for the novel and particularly for the atypical PKC isoenzymes. This simple study on the isoform selectivity of PKC inhibitors demonstrates that virtual screening using protein models can provide useful information for the development of biopharmaceuticals with desired specificity. Despite showing a classification accuracy that is better than random, there is still the possibility of further improvements. However, these would require an alternate approach that focuses on the variability across homologues rather than on their conserved features.

Figure 11.

Prediction of the PKC isoenzyme selectivity of known PKC inhibitors from MDDR. A – three-state binding assignment of good (IC₅₀ <100 nM), weak (100nM < IC₅₀ < 1 μM) and non-binders (IC₅₀ >1 μM) by machine learning. B – number of MDDR compounds predicted to inhibit different PKC isoforms with IC₅₀ <100 nM, C – number of hits returned by the Google search engine (http://www.googlefight.com/) using different PKC isoenzyme inhibitors as the query phrases.

3. 8. Simulation times

Computational procedures were carried out on IBM cluster with 2.0GHz AMD Opteron processors and deploying Linux OS. Figure 12 shows docking times for the programs used in this study. FINDSITE^LHM is the least CPU-expensive procedure with an average docking time of less than 2 min per compound. Q-Dock^LHM requires ~8 min to dock a ligand on average. High-resolution refinement by AMMOS typically uses less than 5 min of CPU time.

Figure 12.

Docking times for FINDSITE^LHM, Q-Dock^LHM and AMMOS. Boxes end at the quartiles Q₁ and Q₃; a horizontal line in a box is the median. “Whiskers” point at the farthest points that are within 3/2 times the interquartile range. Outliers and suspected outliers are presented as solid and blank circles, respectively.

4. DISCUSSION

The increasing interest in kinase inhibitors as novel therapeutics has created a demand for the structural characterization of the human kinase family. Targeting the entire family rather than individual members gives better prospects for developing compounds with improved selectivity ^{106, 107} or, in some cases, inhibitors that are “selectively unselective” i.e. modulate activity of multiple kinase targets associated with the selfsame pathological process ^{88, 108}. Despite progress in protein crystallography and structural genomics efforts that doubled the rate of experimental structure determination ¹⁰⁹, the structural coverage of the kinase family remains poor and unequally distributed ¹¹⁰. Propitiously, the presence of a sufficient number of template structures in the PDB ³⁴ and the high structural conservation of kinase domains make the members of the kinome family perfect targets for template-based structure modeling. A wide range of highly accurate protein models would not only contribute directly to the structure-based drug design ¹¹¹, but also to the initial experimental structure determination of new kinases by molecular replacement techniques ¹¹².

In this study, we constructed reliable three-dimensional models for all kinase sequences identified in the human proteome for use in structure-based drug design. Structure modeling was followed by a detailed functional characterization, starting from the identification of ATP-binding pockets that are the primary target sites for most of the currently available kinase inhibitors ^{64, 89, 113}. Highly accurate protein models and the availability of ligand-bound template structures resulted in precisely annotated binding residues, which constitute a practical dataset to guide further mutational studies. Next, for each kinase family member, we applied fast fingerprint-based virtual screening to rank a collection of >2×10⁶ compounds from the ZINC database ⁴⁹. By selecting the top 10,000 molecules for each kinase, a kinase-focused library of ~30,000 unique compounds was compiled. This collection, representing reasonable chemical coverage of kinase inhibitor space, should improve the efficiency of drug development. In high throughput screens, large combinatorial libraries are frequently supplemented with the target-oriented libraries ^{114, 115}. Recent screening experiments on 41 kinases demonstrated that the overall hit enrichment is significantly higher for a target class focused library compared to generic drug-like compounds ¹¹⁶. Our kinase-focused, 30,000-compound library compiled from the top virtual screening hits may be of practical use for the selection of compounds for high-throughput screens by providing scaffolds with high kinase inhibitory potential.

Docking benchmarks carried out for modeled kinase structures demonstrate that ligand homology modeling often produces approximately correct binding poses, which recover most of the native protein-ligand contacts. These results, nota bene non-trivial, since the distorted binding sites in protein models represent a considerable challenge for many ligand-docking algorithms, are in good agreement with our previous studies ^{38, 39}. We note that over five million distinct models of three-dimensional protein-drug complexes have been constructed; these can be used for rapid binding affinity assessment by any structure-based scoring function.

Our retrospective virtual screening analyses validate the modeled kinase structures as valuable targets in structure-based drug development. Here, we applied a hierarchical virtual screening approach. First, a large collection of compounds was assessed by a fast fingerprint-based approach. Subsequently, the top-ranked fraction of the screening library was submitted to more CPU-expensive ligand homology modeling followed by low-resolution docking/refinement. In the end, lead candidates were re-ranked using structure-based scoring functions. Such a workflow is very common in modern virtual screening protocols that typically consist of a cascade of different filter approaches ¹¹⁷. The least computationally expensive ligand-based techniques applied at the outset of in silico screening allow for a rapid assessment of large compound libraries, with the top fraction of the ranked library enriched with active compounds ^{39, 56, 100}. These pre-filtered subsets are subject to structure-based virtual screening by flexible ligand docking. Predicted binding modes in the target receptor pockets are re-ranked according to the energy of binding estimated from molecular interactions. Finally, the top fraction of the library, typically containing hundreds to thousands molecules, is submitted for experimental validation. Following a protocol of consecutive hierarchical filters, lead candidates that show IC₅₀ values in the micro to nanomolar range have been successfully identified for, e.g., the human aldose reductase ¹¹⁸ and the human carbonic anhydrase ¹¹⁹. Our approach to virtual screening that combines ligand homology modeling and low-resolution docking can be applied to theoretically modeled receptor structures and yields accuracy at least comparable to structure-based virtual screening against high quality X-ray structures using state-of-the-art docking algorithms.

5. CONCLUSIONS

Considering the accelerated pace of genome sequencing and the much slower rate of experimental protein structure determination, it is unlikely that three-dimensional structures will be soon available for all potential drug targets. Therefore, modern drug development at the proteome level must rely on modeled structures provided by state-of-the-art protein structure prediction techniques. In this study, we show that hierarchical virtual screening combining fast fingerprint-based filtering with structure-based ligand homology modeling emerges as a powerful compound prioritization technique applicable to the early stages of proteome-scale drug design projects. By applying this approach to all kinase domains in humans, we have provided the scientific community with a very extensive structural and functional characterization of the human kinome to support the discovery of novel kinase inhibitors.