Targeting Aurora2 Kinase in Oncogenesis: A Structural Bioinformatics Approach to Target Validation and Rational Drug Design¹

Introduction

In cell division, the M phase is composed of mitosis and cytokinesis. Progression through the M phase is dependent on several mitotic kinases, the activities of which are regulated by phosphorylation-dephosphorylation and proteolysis. Of the mitotic kinases, cyclin-dependent kinase 1 is the most prominent cell cycle regulator that orchestrates M-phase activities. However, a number of other mitotic protein kinases that participate in M phase have been identified; they include members of the polo, aurora, and NIMA (never in mitosis A) families and kinases implicated in mitotic checkpoints, mitotic exit, and cytokinesis (1). Aurora kinases are a family of oncogenic S/T kinases³ that localize to the mitotic apparatus (centrosome, poles of the bipolar spindle, or midbody) and regulate completion of centrosome separation, bipolar spindle assembly, and chromosome segregation. Three human homologues have been identified (aurora1, aurora2, and aurora3; Ref. 2), and they all share a highly conserved catalytic domain located in the COOH terminus, but their NH₂ terminal extensions are of variable lengths with no sequence similarity (3). The human aurora kinases are expressed in proliferating cells and are also overexpressed in numerous tumor cell lines of the breast, ovary, prostate, pancreas, and colon (2, 4–6). It has been shown that aurora2 acts as an oncogene and is able to transform both Rat1 fibroblasts and mouse NIH3T3 cells in vitro, and aurora2 transforms NIH3T3 cells grown as tumors in nude mice (3, 4). It has been proposed (2) that overexpression of aurora2 may drive cells into aneuploidy, which in turn can accelerate the loss of tumor suppressor genes and/or amplification of oncogenes, leading to cellular transformation. This is likely to be achieved when those cells that overexpress aurora2 escape mitotic check points, which allows inappropriate activation of proto-oncogenes. Recently, we have shown up-regulation of aurora2 in a number of pancreatic cancer cell lines and have validated this as a target by demonstrating that antisense oligonucleotide treatment leads to a loss of aurora2 activity and increased apoptosis (7). Therefore, aurora2 is a potential target that requires further validation by the rational design of novel small-molecular inhibitors (8) and testing in pancreatic cancer cell lines and animal models for efficacy and potency.

To validate aurora2 as a drugable target in pancreatic cancer, we have undertaken a study to design specific inhibitors of this mitotic S/T kinase. Here we describe a structure-based design approach that uses three-dimensional modeling of aurora1 and aurora2 kinases. Homology modeling of PKs is well validated and documented (9–11). We have used a structural bioinformatic algorithm that uses PSI-BLAST (NCBI), THREADER, 3D-PSSM (three-dimensional position-sensitive scoring matrix), and SAP programs to determine the optimal template for homology modeling of aurora1 and aurora2. The crystal structure of the activated form of bovine cAMP-dependent PK was used as the best template for homology modeling using MD simulations in INSIGHT II. The modeled aurora2 structure was docked with known S/T kinase and aurora2 kinase inhibitors using the binary complex of cAMP-dependent PK-Mn²⁺-AMP-PNP (12, 13). The calculated binding energies from the docking analysis are in agreement with experimental IC₅₀ values obtained from an in vitro kinase assay, which uses histone H1 phosphorylation to assess inhibitor activity.

Materials and Methods

Sequence and Structure Analysis.

The aurora2 (ARK1; residues 98–403) and aurora1 (ARK2; residues 35–344) kinase domain sequences were used as probes to search a nonredundant database of sequences using PSI-BLAST (NCBI). Top-ranked sequences for which the three-dimensional structures of S/T kinase domains are available were the porcine heart cAMP-dependent PK (1CDK; Ref. 12), recombinant mouse cAMP-dependent PK (1APM; Ref. 14), and Caenorhabditis elegans twitchin kinase (1KOA; Ref. 15). The aurora1 and aurora2 domain sequences were analyzed using the programs THREADER (16) and 3D-PSSM, (17), which compare primary sequences with all of the known three-dimensional structures in the Brookhaven Protein Data Bank. The output is composed of the optimally aligned, lowest-energy, three-dimensional structures that are similar to the aurora kinases. The top-ranked structures were bovine cAMP-dependent kinase (1CDK), murine cAMP-dependent kinase (1APM) and twitchin kinase (1KOA) confirming the PSI-BLAST search. These three tertiary structures provide the three-dimensional templates for the homology modeling of aurora1 and aurora2. The crystal structure coordinates for the above S/T kinase domains were obtained from the Protein Data Bank (18). These domains were pairwise superimposed onto each other using the program SAP (19). The structural alignments from SAP were fine-tuned manually to better match residues within the regular secondary structural elements. The three manually aligned S/T kinase domain sequences with their respective secondary structures were viewed in Clustal X (20).

Homology Modeling and Refinement.

The modeling software used was INSIGHT II (version 2000, Accelrys Inc.) running on a Silicon Graphics Indigo2 workstation (Silicon Graphics Inc.) under the UNIX operating system. The crystal structure coordinates of the ternary complex of cAMP-dependent kinase (1CDK; Ref. 12) were retrieved from the Brookhaven Protein Data Bank. The catalytic domain served as the template for aurora1 and aurora2 modeling. The inhibitory peptide PKI (5–24) and the solvent molecules were deleted from the coordinate set before modeling. The two Mn²⁺ ions in the active-site pocket were retained and replaced by Mg²⁺ ions. The model building procedure for aurora1 and aurora2 entailed a deletion of residues ²⁸²GNLKD²⁸⁶ from the COOH-terminal end of cAMP-dependent PK and a deletion of the last 30 residues from the COOH terminus of cAMP-dependent PK (residues highlighted in blue in Fig. 1). Ser²⁸⁴ and Leu²²⁸ were inserted using the conformation of Lys²⁰⁴ of twitchin kinase (1KOA). All of the other side chains except for the identical residues were mutated to those of aurora1 using the Homology modeler and Biopolymer modules in Insight II.⁴ This model served as a template for the homology modeling (12, 13) of aurora2 kinase. The main differences between aurora1 and aurora2 are two insertions between Lys¹²⁴-Lys¹²⁵ and Ser³⁸⁸-Lys³⁸⁹. The peptide bonds between these residues were broken, and SGTPDIL and RVL residues were inserted in aurora1 consecutively without altering the geometry of the rest of the sequence. The conformation of the inserted fragments was obtained from the fragment database of low-energy conformations. Furthermore, all of the amino acid residues except for the identical residues were mutated to those of aurora2. The mutated side chains were manually positioned to minimize steric hindrance with adjacent residues. In general, insertions and deletions were found in loop regions connecting regular secondary structures and, hence, did not alter the hydrophobic core of the structure.

After the model building processes were complete, a series of minimizations were performed to relax the structure. Before model refinement, all of the close contacts caused by the mutation of side chains were fixed by manually rotating the χ₁ and χ₂ torsion angles and keeping the side chain torsion angles of the conserved residues in their original conformation. Hydrogen atoms were supplied by Biopolymer with CFF parameters (21), solvated by simulating with the distance dielectric constant, ϵ = 4rij (22, 23) and refined by the steepest descent and conjugate gradient minimization until convergence was reached, while restraining the position of the heavy atoms. Finally, the entire system was subjected to energy minimization using the 3000-step steepest descent followed by conjugate gradient minimization until an energy gradient of less than 0.001 kcal/mol/Å was achieved and from which, constraints were gradually removed.

Model Evaluation.

The final aurora2 model was examined using profile-3D (24). The profile-3D and three-dimensional-one-dimensional score plots of the model were positive over the entire length of protein in a moving-window scan to the template structure. Additionally, PROCHECK (25) was used to verify the correct geometry of the dihedral angles and the handedness of the model-built structure. For the PROCHECK statistics, an overall G factor of −0.31, hydrogen bond energy of 1.6, and only 0.29 bad contacts per 100 residues were observed, which is consistent with a good quality structure comparable with the crystal structure. The overall fold of the homology model (Fig. 3A) was found to be very similar to that of crystal structure template 1CDK, which was used for additional energy refinement, MD simulations, and docking studies of other S/T kinase inhibitors.

MD Simulations.

The energy-minimized aurora2 structure was used as starting model for MD simulations (26, 27), which were performed in the canonical ensemble (NVT) at 300°K using the CFF force field implemented in Discover_3 (version 2.9.5).⁵ Dynamics were equilibrated for 10 ps with time steps of 1 fs and continued for 10-ps simulations. The nonbonded cutoff distance of 8 Å and a distance-dependent dielectric constant (ϵ = 4rij) for water were used to simulate the aqueous media. All of the bonds to hydrogen were constrained. Dynamic trajectories were recorded every 0.5 ps for analysis. Individual simulations from the point of a stable trajectory generated time-averaged structures. The resulting low-energy structure was extracted, and energy was minimized to 0.001 kcal/mol/Å. To examine the conformational changes that occur during MD, the rms deviations were calculated from trajectories at 0.5-ps intervals and compared with the Cα backbone of cAMP-dependent PK (see Fig. 3B). The rms deviation for the two superimposed structures was 0.42 Å. Furthermore, the rms deviations were calculated for the protein backbone (0.37 Å) and the active-site pocket (0.41 Å) and were compared with crystal structure before the docking experiments. The resulting aurora2 structure served as the starting model for docking studies.

SA Docking.

To find sterically reasonable binding geometries and to explore the interactions of the proposed ligand recognition pocket, affinity docking with SA was performed in INSIGHT II (28, 29). The structures of ligands used for docking were from five crystal structure complexes of cAMP-dependent PK bound to AMP-PNP, (12), staurosporine (30, 31), H-89, H-7, H-8, (32, 33), and structures that were empirically built and energy minimized [KN-93 (34), ML-7 (35), and 6,7-dimethoxyquinazoline (8, 36); Fig. 2] in INSIGHT II. Partial charges were assigned to the ligands by the Gasteiger method defined within INSIGHT II.⁵ Systematic conformational searches were performed on each of the minimized ligands using 10-ps MD simulations at 300°K. For docking with AMP-PNP, the position of the ATP analogue was retained from its crystal structure with 1CDK, in which the adenine base served as a template for field-fit alignments with the indolocarbazole, fused biphenyl, and quinazoline moieties. The ATP analogue was then removed from the field-fit alignment, and each of the other ligands was docked into the active-site pocket with a position and orientation that were similar to those of AMP-PNP. The heavy atoms from the ATP analogue were used as sphere centers in the input to the docking procedure. To clarify the orientation of these inhibitors in the active-site pocket, the electrostatic potential at the van der Waals surface was determined using solvent surface calculations. Orientations with the lowest intermolecular potential energy were calculated. To explore the interaction of the ligands in the recognition pockets, SA docking of complex formations was carried out without constraints, to allow each of the protein-ligand complex systems to evolve freely. To explore the effects of solvation implicitly, a distance-dependent dielectric constant (ϵ = 4rij) was used (22, 23). The nonbonded cell multipole method was used for SA docking with input energy parameters. Docking simulations were performed at 500°K with 100 fs/stage (total of 50 stages), quenching the system to a final temperature of 300°K. The whole complex structure was energy minimized using 1000 steps. This provided 10 structures from SA docking, and their generated conformers were clustered according to rms deviation. The lowest global structure obtained was used for computing intermolecular binding energies (37). Furthermore, we have validated the robustness of the affinity-docking methodology (27, 38) by comparative docking of an ATP analogue, AMP-PNP, with 1CDK and aurora2 kinase. The binding mode of these complexes are shown in Fig. 4, A and B, respectively.

In Vitro Aurora Kinase Assay.

Aurora2 immunoprecipitation kinase assays were performed as described previously (6). Briefly, cells were first homogenized in lysis buffer [150 mm NaCl, 50 mm HEPES (pH 7.2), 1 mm EDTA, 1 mm EGTA, 1 mm DTT, 0.1% Tween 20, 0.1 mm phenylmethylsulfonylfluoride, 2.5 μg/ml leupeptin, 0.1 mm sodium orthovanadate (Sigma Chemicals, St. Louis, MO)] at 4°C. Lysates were then centrifuged at 10,000 × g for 10 min. Protein content was determined by the BCA protein assay (Pierce, Rockford, IL), and 200 μg was used for each sample. The supernatants were immunoprecipitated for 12 h at 4°C with protein A-agarose beads precoated with saturating amounts of aurora kinase-2 polyclonal antibody. Immunoprecipitated proteins on beads were washed twice with 1 ml of lysis buffer and twice with kinase buffer [50 mm HEPES (pH 7.0), 10 mm MgCl₂, 5 mm MnCl₂, and 1 mm DTT]. The beads were then resuspended in 40 μl of kinase buffer containing 10 μm ATP, 5 μCi of [γ-³²P]ATP (6000 Ci/mmol; 1 Ci = 37 GBq; Amersham Corp., Arlington Heights, IL), and 4 μg of the histone H1 protein substrate. The samples were then incubated for 30 min at 30°C with occasional mixing. For dose-response studies to determine IC₅₀, serial dilutions of the inhibitors were used (10–200 μm). The samples were then boiled in polyacrylamide gel sample buffer containing SDS and separated by electrophoresis. Phosphorylated proteins were quantified after exposure to autoradiographic film (Labscientific, Inc., Livingston, NJ) by densitometry using ImageQuant version 5.1 (Molecular Dynamics Computing Densitometer, Sunnyvale, CA). The IC₅₀ values (Table 1) were calculated from the phosphorylation profile of the substrate histone H1 protein.

Results

Aurora Kinases Possess a Conserved yet Distinct S/T Kinase Catalytic Domain

The sequence identity and similarity between aurora1 and aurora2 kinases and cAMP-dependent PK are 30 and 60%, respectively, and that between aurora1 and aurora2 kinases and twitchin kinase is 17 and 54%, respectively. The aurora1 and aurora2 domain sequences were aligned with respect to the structural alignments obtained in Clustal X. Fig. 1 shows the secondary structural elements of the aligned kinase domains. The critical catalytic residues involved in the transfer of the γ-phosphoryl group of ATP to the substrates are highly or absolutely conserved between the aurora kinases, cAMP-dependent PK, and twitchin kinase, and these residues are highlighted in Fig. 1. In the prototype cAMP-dependent PK catalytic domain, they are ⁵⁰GXGXXGX⁵⁷, ⁷²K, ⁹¹E, ¹²¹E, ¹⁶⁶D, ¹⁶⁸K, ¹⁷⁰E, ¹⁷¹N, ¹⁸⁴DFG, ¹⁹⁷T (autophosphorylation site), ²⁰⁶APE, ²²⁰D, and ²⁸⁰R (12). The main differences between cAMP-dependent PK and the aurora kinases are: (a) the glycine-rich nucleotide binding motif ⁵⁰GTGSFGRV⁵⁷ is changed to ¹⁴⁰GKGKFGNV¹⁴⁷; (b) ²⁰⁶APE is changed to ²⁹⁷PPE²⁹⁹; (c) ²⁸⁴S and ²²⁸L in aurora kinases are inserted; (d) residues GNLKD are deleted from the COOH-terminal end of cAMP-dependent PK; and (e) the last 30 residues are deleted from the COOH terminus of cAMP-dependent PK. These results indicate that the aurora kinases possess a conserved S/T kinase catalytic domain similar to that of the cAMP-dependent PKs. However, the changed glycine-rich nucleotide-binding sequence motif, V¹²³A²¹³, E¹²⁷T²¹⁷, insertion of Ser²⁸⁴ in aurora2 and Leu²²⁸ in aurora1 (between Ser²⁸³ and Arg²⁸⁵) and A²⁰⁶P²⁹⁷ in the active site provide important structural differences between these two kinases that may explain their distinct substrate specificities.

Each S/T kinase structure is composed of a small NH₂-terminal β-sheet domain and a large COOH-terminal α-helical domain, with the catalytic cleft between the two domains at which the substrate and ATP bind in the presence of Mg²⁺ ions (12). The crystal structure determined for bovine cAMP-dependent PK is phosphorylated in the activation loop and is in complex with AMP-PNP (Fig. 2) and the pseudo-substrate peptide PKI (5–24). This catalytically competent enzyme is in a state immediately before phosphoryl transfer, with the cleft between the NH₂- and COOH-terminal domains in a closed conformation. The tertiary structure modeling of aurora1 and aurora2 on the catalytically competent cAMP-dependent PK structure has defined pertinent structural features in the aurora kinases and has guided our structure-based rational design process. Fig. 3A shows the model-built structure of aurora2. Secondary structural elements are indicated as red cylinders (α-helix), yellow arrows (β-sheet), green ribbons (coil), and blue ribbons (turns). The Cα backbone superimposition of the final minimized structure of aurora2 (blue) on cAMP-dependent PK (1CDK; red) is shown with a rms deviation of 0.42 Å (Fig. 3B). Because of the two COOH-terminal deletions in the aurora kinases compared with that of cAMP-dependent PK, the aurora kinases have evolved a more compact catalytic kinase core domain that may affect their substrate specificity and activity.

Docking of S/T Kinase Inhibitors Coupled to an in Vitro Kinase Assay Defines Distinct Chemical Moieties as Building Blocks for Potent Inhibitors of Aurora2

Aurora2-AMP-PNP Complex.

Affinity docking experiments indicate that AMP-PNP in complex with aurora2 exhibits similar conformational orientations to that of the crystal structure of cAMP-dependent PK bound to AMP-PNP (12). This comparative binding mode provides validation of affinity docking and a rationale for docking simulations of other S/T kinase inhibitors. Because of the change in sequence motif of the glycine-rich loop and residues interacting with the ATP analogue, AMP-PNP binds with an affinity in the ATP-binding pocket of aurora2 that is lower than that with 1CDK (Table 1). The crystal structure of the 1CDK-AMP-PNP complex (12) revealed that the high binding affinity of the ATP analogue in the nanomolar range is caused by a number of interactions, including polar, β,γ-phosphoryl coordination to Mn²⁺ ions, and hydrogen bonds with specific enzyme residues. From our docking and MD simulations, we observed a network of hydrogen bonding interactions with aurora2, which, in turn, stabilized the docked structure. The adenine base, its 6-amino group, and ring N1 atoms interact with Glu²¹¹ and Ala²¹³, which have strong hydrogen bonding interactions: Glu²¹¹-C⋕O… H₂N of adenine (1.97 Å, energy −2.34 kcal/mol) and Ala²¹³-NH… N of ring N1 (2.26 Å, energy −2.65 kcal/mol), respectively. Weaker hydrogen bonding interactions for AMP-PNP with Glu¹²¹ (2.90 Å) and Val¹²³ (3.20 Å) were observed in the 1CDK crystal structure complex. In comparative docking experiments, the docking of AMP-PNP with 1CDK revealed that the conformation and binding mode of AMP-PNP shows high similarity to that of the crystal structure (Fig. 4B). Another important finding from docking simulations suggests that the γ-phosphoryl group seems to play a significant role in hydrogen bonding interactions with the −NH atoms of Phe¹⁴⁴ and Gly¹⁴⁵ in aurora2. The 1CDK crystal structure complex shows that the 2′ hydroxyl group of the ribose moiety has potential hydrogen bonding interactions with Glu¹²⁷ and Glu¹⁷⁰. However, in the aurora2-AMP-PNP complex, the 2′OH group of the ribose moiety has a H-bonding interaction with Thr²¹⁷ (Thr²¹⁷-O… H of 2′OH-group, 2.40 Å). The conformation of AMP-PNP within the active site of aurora2 is maintained by seven hydrogen bonds formed arising from the ATP-binding site and the Gly/Lys pocket (Fig. 4). These hydrogen bonding interactions were computed from the final MD trajectories of the docked complex structure. The absence of the Glu²⁶⁰ hydrogen bonding interaction with the ribose moiety of aurora2 is attributable to the orientation of the β,γ-phosphoryl groups in the Gly/Lys pocket that prevent a strong electrostatic interaction with Lys¹⁶². The β- and γ-phosphoryl groups are also involved in close contacts with Lys¹⁴³, Asn²⁶¹, and Asp²⁷⁴ of the active site. The absence of critical H-bonding interactions with Lys¹⁶² and Glu²⁶⁰ presumably leads to an affinity binding of AMP-PNP (−69.3 kcal/mol) with the active-site residues of aurora2 that is lower than that with 1CDK (Table 1). These sequence and structural differences between cAMP-dependent PK and aurora2 kinase provide guidelines to better define the active-site pocket and to design novel and more specific aurora2 kinase inhibitors.

Aurora2-Staurosporine Complex.

In the case of affinity SA docking simulations on staurosporine (Fig. 2), the position and orientation within the active-site pocket of aurora2 were identical to those of the cAMP-dependent PK-staurosporine complex (1STC; Ref. 30). In particular, the hydrogen bonding interactions computed from affinity were identical to those in the crystal structure of the cAMP-dependent PK-staurosporine complex (Fig. 5). The amino acid residues that recognize and interact with staurosporine are Gly¹⁴⁰, Val¹⁴⁷, Glu¹⁸¹, Glu²¹¹, Ala²¹³, Thr²¹⁷, Asn²⁶¹, and Asp²⁷⁴. A strong intermolecular hydrogen bond to Ala²¹³ −NH with the lactam carbonyl group (2.03 Å) of staurosporine was observed in all of the structures obtained from MD simulations. However, Glu²¹¹-C⋕O… HN of the lactam amide hydrogen bonding interaction, present in the cAMP-dependent PK-staurosporine complex, was not observed in the initial docked aurora2-staurosporine complex. This is most likely attributable to the indolocarbazole and pyranosyl moieties of staurosporine, which induced conformational changes in the neighboring active site residues of aurora2. The staurosporine-induced conformational changes were investigated by calculating the rms between the initial and final complex structures obtained from MD simulations. The 0.23-Å difference observed in the active-site pocket of aurora2 on staurosporine binding is shown in Table 2. An analysis of the low-energy trajectory from the final run of simulations reveals that the amide proton forms an H-bond interaction with Glu²¹¹ (2.72 Å). It is interesting to note that the indolocarbazole moiety exhibits a more stable orientation in all of the structures generated during MD simulations. The phenyl ring syn to the methoxy group is positioned in the ATP-binding pocket and is in close contact with Glu¹⁸¹ and Asp²⁷⁴ residues and are within 2.50–3.00 Å, whereas the phenyl ring anti to methoxy group is oriented into the Gly-Lys pocket. The pyranosyl methylamino group is positioned within H-bonding distance to Thr²¹⁷ (1.97 Å) and Glu²⁶⁰ (3.32 Å). On the basis of these observations and experimental data (Table 1), staurosporine (a nanomolar PK inhibitor) binds to aurora2 with a high binding energy of −91.6 kcal/mol (see Table 1).

Aurora2-H-series Complexes.

We have analyzed the mode of binding of three isoquinoline H-series PK inhibitors, H-89, H-8, and H-7 (Fig. 2), with aurora2 kinase (32, 33). The conformations of these isoquinolines were obtained from the crystal structure complexes of cAMP-dependent PKs and were subjected to affinity docking experiments without altering their crystallographic geometries. AMP-PNP served as a template for SA docking, and an analysis of dynamic trajectories revealed that the isoquinoline moiety occupied the ATP binding site and is positioned similarly to that of the adenine base of AMP-PNP. The N-alkyl chain of H-8 and H-89 extended into the β,γ-phosphoryl site. In all of the H-series isoquinolines, we observed that the nitrogen ring interacts with Ala²¹³, which is located close to Glu²¹¹ at the hinge region and is in the purine binding site of AMP-PNP. Although the H-89 isoquinoline structure is different from that of the AMP-PNP, it has similar structural elements, including an isoquinoline nitrogen atom at position 13, two −NH groups at positions 4 and 17, and a sulfonyl group at position 1. These play an important role in the recognition of the active-site pocket of aurora2 and represent another class of S/T kinase inhibitors with an IC₅₀ of 107 μm (Table 1). A number of SA docking simulations with different initial configurations identified a binding mode with a binding energy (−103.5 kcal/mol) substantially higher than that observed for AMP-PNP, staurosporine, and other isoquinoline structures. In this model, the N13 atom is positioned such that it H-bonds with the Ala²¹³ −NH group (−N… HN-Ala²¹³, 2.80 Å), which mimics the 1CDK Val¹²³ amide H-bonding interaction (Fig. 6). Unlike the β,γ-phosphoryl group, the bromophenyl moiety is located deep in the Gly/Lys pocket and exhibits close hydrophobic interactions (3–3.50 Å) with Phe¹⁴⁴ and Gly¹⁴⁵. The carbonyl group of Asn²⁶¹ in aurora2 forms a strong hydrogen bonding interaction with the N17 amide proton, and Glu²⁶⁰ is also in close contact with the same amide proton. The isoquinoline ring and sulfonamide moiety make a number of close contacts with Gly¹⁴⁰, Gly¹⁴⁵, and Val¹⁴⁷. Moreover, the styrene moiety is also in close contact with Lys¹⁴¹ and Gly¹⁴². These interactions determine a stable binding mode observed in several dynamic trajectories analyzed within ±10 kcal/mol of that of the low-energy global structure generated during MD simulations.

For the H-8 and H-7 compounds (data not shown), the docking is based on the aurora2-H-89 complex. The N13 isoquinoline ring atom involved in H-bonding interactions with the Ala²¹³ amide group is within a distance of 1.95–2.45 Å. The replacement of the bromophenyl styrene moiety of H-89 with the N-alkyl side chain and piprazine group in the case of H-8 and H-7 leads to less potent inhibitors of aurora2. Although the isoquinoline and sulfonyl moieties exhibit structural similarity, the H-series of compounds lacks a number of van der Waals and hydrophobic contacts with the enzyme. Such interactions play an important role for high binding affinity, as observed in the crystal structures of cAMP-dependent PK in complex with inhibitors. As shown in Fig. 6, the simulations demonstrate those structural changes that alter the shape of the sulfonyl substituents, leading to significant changes in binding energies (Table 1). Thus, the −NH of the N-methyl group of H-8, oriented to the Asp²⁷⁴ residue and its methyl function, are positioned within 3.5–4 Å. The sulfonyl substituent had only one contact with Val^l47. Whereas the piprazine moiety is positioned within the ribose binding site of AMP-PNP, its −NH group is 3.5 Å away from Lys¹⁴¹. These results suggest that although H-8 and H-7 compounds bind to aurora2, the higher affinity of H-89 can be explained by its strong hydrogen bonding and hydrophobic interactions with both the ATP-binding site and the residues within the Gly/Lys pocket.

Aurora2-6,7-Dimethoxyquinazoline Complex.

Although quinazoline-containing compounds have been reported to be very potent and selective EGFR receptor tyrosine kinase inhibitors (39, 40), recent data suggest that these compounds also bind to S/T kinases (8). The 6,7-dimethoxyquinazoline which binds to aurora2 with an IC₅₀ value of 0.00785 μm (8), represents another class of high-affinity aurora2 inhibitors. The aurora2-H-89 complex provided the basis for docking the quinazoline ring and, subsequently, the H-89 molecule was removed. The entire aurora2-6,7-dimethoxyquinazoline complex was optimized by several cycles of SA docking simulations and minimization using the CFF force field. From SA docking simulations, it was observed that the quinazoline moiety was positioned in an orientation that was similar to that of the adenine base of AMP-PNP and was anchored deep in the ATP binding site by forming three hydrogen bonds (Fig. 7). The N1 atom of quinazoline and N3 atom of the pyrimidine rings are involved in potential hydrogen bonding interactions with Ala²¹³ (Ala²¹³-NH… N1, 2.78 Å, energy −1.61 kcal/mol) and Asp²⁷⁴ (Asp²⁷⁴-NH… N3, 2.05 Å, energy −1.29 kcal/mol). Additionally, the −C⋕O of Asp²⁷⁴ is involved in hydrogen bonding interactions with 2-(N-benzoylamino) −NH group (Fig. 7). The oxygen atom of the 7-methoxy substituent is also positioned within hydrogen bonding distance with the backbone amide of Ala²¹³. Glu²¹¹ has strong hydrogen bonding interactions with AMP-PNP and staurosporine, but no such an interaction was observed with the quinazoline compound. This is attributable to the lack of a donating functional group and strong steric interaction of methoxy group with Glu²¹¹. The 2-(N-benzoylamino) group had a similar position and orientation to that of the γ-phosphoryl group of AMP-PNP or the bromophenyl group of H-89 and interacts with Phe¹⁴⁴ and Gly¹⁴⁵, which in turn interacts with Lys¹⁴³ located 3.50 Å from the carbonyl group. In the final model, we observed that the 5-aminopyrimidine moiety and the corresponding H-89 sulfonyl moiety positioned similarly with in the active site. As shown by the AMP-PNP model, the similar manner in which H-89 and quinazoline (Fig. 8) bind indicates that the polar, nonpolar, and hydrogen bonding interactions are essential for the high activity observed for these compounds. Such a large number of polar, nonpolar, and hydrogen bonding contacts with the active-site residues are observed for 6,7-dimethoxyquinazoline, which exhibits a high binding energy (−104.9 kcal/mol) compared with that of H-89 (Table 1). These results suggest that the rigid substituents of the β- and γ-phosphoryl positions are well tolerated.

Aurora2 in Complex with KN-93 and ML-7.

The binding modes of KN-93 and ML-7 (data not shown) were also explored using the above models. ML-7 was positioned and oriented similarly to H-7. The 4-methoxyphenyl sulfonyl amide moiety of KN-93 was found to occupy the ATP binding site of aurora2. The interaction of the chlorophenylstyrene group with Asp²⁷⁴ and Asn²⁶¹ and its binding mode is dissimilar to that observed for the inhibitor class of H-89 and H-8 bound to aurora2. Interestingly, KN-93 did not exhibit any hydrogen bonding interactions. Specifically, the presence of the methyl (amino) methyl and hydroxyethyl groups had significant steric clashes with Gly¹⁴⁰, Val¹⁴⁷, and Asp²⁷⁴, presumably because of its rigid conformation.

Discussion

The development of potent and selective nucleotide analogue-based inhibitors that interact with individual PKs at their ATP binding site are being evaluated in various stages of clinical trials (41). Given the proof of concept that the inhibition of dysregulated PKs in human cancers is potentially achievable and could provide clinical efficacy, we have targeted human aurora2 kinase in pancreatic cancer. An iterative structure-based small-molecule drug design approach in combination with an in vitro kinase assay has been developed to test the potency and efficacy of novel aurora2 kinase inhibitors.

The tertiary structure models of aurora1 and aurora2 were built using the catalytic domain of PKA complexed with a 20-amino acid substrate analogue inhibitor and Mn²⁺-ATP (12). The cAMP-dependent PK structure revealed a two-domain PK with a deep cleft between the domains. Mn²⁺-ATP and a portion of the inhibitor peptide occupy the active site cleft. The NH₂-terminal smaller domain is associated with nucleotide binding and its largely antiparallel β-sheet structure constitutes a nucleotide binding motif. The COOH-terminal larger domain is predominantly α-helical with a single β-sheet at the domain interface. This domain is primarily involved in peptide binding and catalysis. Most of the invariant amino acids in this conserved catalytic core are clustered at the sites of nucleotide binding and catalysis. The tertiary structure of the aurora2 model built on 1CDK has significant identity and similarity indicating a conserved catalytic core. The main differences in the active site of aurora2 and cAMP-dependent PK are the changed glycine-rich nucleotide-binding motif ⁵⁰GKGKFGNV⁵⁷, which should be sufficient to provide distinct substrate specificities and deletion of the last 30 residues from the COOH terminus of cAMP-dependent PK (Fig. 1) and to provide for a more compact aurora2 catalytic kinase core domain that may be relevant to its functional localization to mitotic spindles.

A number of crystal structures of cAMP-dependent PK in complex with ATP analogues and pseudo-substrates (42, 43) have provided details as to the mechanisms of inactivation by modulation of one or more of the four conserved structural elements; these are: (a) inhibition of the ATP binding pocket; (b) distortion of the glycine-rich loop; (c) alteration of the position of α-helix C; and (d) alteration of the conformation of the phosphorylation-dependent activation segment. The crystal structure determined for bovine cAMP-dependent PK is phosphorylated in the activation loop and is in complex with AMP-PNP and the pseudo-substrate peptide PKI (5–24). This catalytically competent enzyme is in a state immediately before phosphoryl transfer, with the cleft between the NH₂- and COOH-terminal domains in a closed conformation. The tertiary structure modeling of aurora1 and aurora2 on the catalytically competent cAMP-dependent PK structure has defined pertinent structural features in the aurora kinases that has guided the rational design of novel chemical entities.

To gain insight into the binding mode of nucleoside analogue-based inhibitors, we have docked known S/T kinase inhibitors. The potency of these docked inhibitors has been evaluated in an in vitro kinase assay. The complex structure of AMP-PNP with cAMP-dependent PK and aurora2 from SA docking simulations reveals a network of hydrogen bonding interactions. The sequence differences in the active-site pocket between the two kinases explains the lower binding energy of the AMP-PNP-aurora2 complex. This is attributable to the weak hydrogen bonding interactions of the adenine base with Glu²¹¹, Ala²¹³, and Glu¹²⁷Thr²¹⁷ mutation in aurora2, the absence of a hydrogen bonding interaction of the 2′OH group of ribose with Lys¹⁶² and Glu²⁶⁰ and the orientation of the β,γ-phosphoryl groups in the Gly/Lys pocket (Fig. 4). However, in the case of staurosporine, the position and orientation in the active site was identical to that of the cAMP-dependent PK-staurosporine complex (1STC; Ref. 30). In particular, the hydrogen bonding interactions computed from affinity were identical with the mostly conserved active-site residues (Fig. 5) and accommodated staurosporine by minor conformational changes during MD simulations (Table 2).

In the case of the H-series of compounds (H-89, H-8, and H-7) the isoquinoline moiety occupied the ATP binding site and is positioned similarly to that of the adenine base of AMP-PNP. The N-alkyl chain of H-89 extended into the β,γ-phosphoryl site and the isoquinoline nitrogen ring interacts with Ala²¹³. Although the H-89 isoquinoline structure is different from that of the AMP-PNP, it has similar structural elements including an isoquinoline nitrogen atom at position 13 which is involved in a hydrogen bonding interaction with Ala²¹³ (Val¹²³ in 1CDK). The bromophenyl moiety is located deep in the Gly/Lys pocket and exhibits a hydrophobic interaction with Phe¹⁴⁴ (Fig. 6). Furthermore, the strong hydrogen bonding and hydrophobic interactions with both the ATP and Gly/Lys pocket supports the calculated binding energy (Table 1). The aurora2-H-89 complex was used to model 6,7-dimethoxyquinazoline, and the quinazoline moiety is positioned similar to that of the adenine base of AMP-PNP and anchored deep in the ATP binding site (Fig. 7). The 2-(N-benzoylamino) group had similar positions and orientations to those of the γ-phosphoryl group of AMP-PNP and interacts with Gly¹⁴⁵, which in turn is in close proximity to Phe¹⁴⁴. The N1 atom of quinazoline and N3 atom of pyrimidine ring exhibits a strong hydrogen bonding interaction with Ala¹²³ and Asp²⁷⁴. In quinazoline, the 2-substituted amide hydrogen is in close contact with Gly¹⁴⁵ and its 5-substituted amide hydrogen has strong interactions with Gly¹⁴⁰ and Val¹⁴⁷. Furthermore, the aminopyrimidine and benzoly carbonyl groups are positioned close to Asp²⁷⁴. These interactions explain its high binding energy (−104.9 kcal/mol) when compared with that of H-89. Compounds KN-93 and ML-7 (data not shown) essentially had poor interactions with aurora2, mainly because of their rigid conformations and lack of hydrogen bonding interactions. The 4-chlorophenyl-2-propenyl-(methylamino)methylphenyl substituent of KN-93 at the 2-position is close to the hydrophilic pocket and appears to be sterically crowded. The 4-position projects toward the Gly/Lys pocket in the region occupied by the bromostyrene moiety of H-89. Taken together, these results suggest that SA docking complexes of aurora2 with H-89 and 6,7-dimethoxyquinazolines furnish a number of guidelines for analogue design.

For the docked structures, the calculated binding energies from the highest to the lowest are 6,7-dimethoxyquinazoline > H-89 > staurosporine > H-8 > AMP-PNP > KN-93 > ML-7 > H-7. For the in vitro kinase assay the IC₅₀ values from the highest to the lowest are 6,7-dimethoxyquinazoline > staurosporine > H-89 > ML-7 > KN-93 (Table 1). These results qualitatively validate the modeling and docking studies described for this series of S/T kinase inhibitors. Molecular modeling was used to evaluate and select potential substituents for the isoquinoline and quinazoline templates. A de novo small fragment and available chemical database library using LUDI⁴ (44) and FlexX (45) virtual docking computational procedures were performed using SYBYL (Tripos, St. Louis, MO) to generate novel compounds specific to aurora2. The high-score compounds from FlexX having a quinazoline moiety were selected, and are currently being synthesized and screened, for aurora2 kinase inhibitory activity.

Recently, the crystal structure of the human aurora 2 kinase (46) was published (Brookhaven Protein Data Bank entry 1MUO on hold). The structure determination was based on molecular replacement using an aurora2 kinase homology model, which was based on 1CDK. This essentially validates the model for docking studies undertaken in this report. The main conclusions from the crystal structure were that the homology model is consistent with an overall kinase fold and active site except for a lack of electron density of the activation loop, which includes a conserved tryptophan. Although the coordinates of the crystal structure of aurora2 are not available, the binding mode of the docked AMP-PNP in the ATP binding site are similar in 1CDK and the modeled aurora2.

In conclusion, we have developed a structural bioinformatic approach that has facilitated the tertiary structure modeling of aurora2, studied the binding mode of ATP analogue and inhibitors, and evaluated these compounds using an in vitro kinase assay. A structural analysis of the docked ATP analogue and S/T kinase inhibitors has provided guidelines for the rational design of novel aurora2 kinase inhibitors. Furthermore, using the available chemical database and FlexX, we have identified a series of novel compounds that are currently been synthesized for in vitro and in vivo testing.