Computational Biology and Chemistry 48 (2014) 1–13
Contents lists available at ScienceDirect
Computational Biology and Chemistry journal homepage: www.elsevier.com/locate/compbiolchem
Research Article
Targeting the Akt1 allosteric site to identify novel scaffolds through virtual screening Oya Gursoy Yilmaz, Elif Ozkirimli Olmez, Kutlu O. Ulgen ∗ Bogazici University, Department of Chemical Engineering, 34342 Istanbul, Turkey
a r t i c l e
i n f o
Article history: Received 23 August 2013 Received in revised form 20 October 2013 Accepted 21 October 2013 Keywords: Virtual screening Molecular docking Induced fit docking Pharmacophore Allosteric Akt inhibitor Allosteric kinase inhibitor
a b s t r a c t Preclinical data and tumor specimen studies report that AKT kinases are related to many human cancers. Therefore, identification and development of small molecule inhibitors targeting AKT and its signaling pathway can be therapeutic in treatment of cancer. Numerous studies report inhibitors that target the ATP-binding pocket in the kinase domains, but the similarity of this site, within the kinase family makes selectivity a major problem. The sequence identity amongst PH domains is significantly lower than that in kinase domains and developing more selective inhibitors is possible if PH domain is targeted. This in silico screening study is the first time report toward the identification of potential allosteric inhibitors expected to bind the cavity between kinase and PH domains of Akt1. Structural information of Akt1 was used to develop structure-based pharmacophore models comprising hydrophobic, acceptor, donor and ring features. The 3D structural information of previously identified allosteric Akt inhibitors obtained from literature was employed to develop a ligand-based pharmacophore model. Database was generated with drug like subset of ZINC and screening was performed based on 3D similarity to the selected pharmacophore hypotheses. Binding modes and affinities of the ligands were predicted by Glide software. Top scoring hits were further analyzed considering 2D similarity between the compounds, interactions with Akt1, fitness to pharmacophore models, ADME, druglikeness criteria and Induced-Fit docking. Using virtual screening methodologies, derivatives of 3-methyl-xanthine, quinoline-4-carboxamide and 2-[4-(cyclohexa-1,3-dien-1-yl)-1H-pyrazol-3-yl]phenol were proposed as potential leads for allosteric inhibition of Akt1. © 2013 Elsevier Ltd. All rights reserved.
1. Introduction The AGC kinase group constitutes 12% of the human kinome and is one of the most evolutionarily conserved groups (Arencibia et al., 2013). Each family of AGC kinases has a physiological role and is involved in human diseases, therefore is a potential drug target. Akt (also named as protein kinase B, PKB), a phospholipid bindingserine/threonine kinase, belongs to AGC superfamily of human kinome and is a key mediator of the PI3K cell survival signaling pathway (Calleja et al., 2009a). Activation through deregulation, amplification and rearrangement of the phosphatidylinositol 3-kinase (PI3K)/Akt pathway occur frequently in many human cancers and is associated with tumor growth, increased metastasis, and resistance to therapy (Hennessy et al., 2005). Consequently, mutations of its components make either itself or its major mediators such as Akt appealing as a therapeutic target for cancer researchers (Liu et al., 2009). The three isozymes of human Akt (Akt1, Akt2,
∗ Corresponding author. Tel.: +90 212 359 6869; fax: +90 212 287 2460. E-mail addresses:
[email protected] (O.G. Yilmaz),
[email protected] (E.O. Olmez),
[email protected] (K.O. Ulgen). 1476-9271/$ – see front matter © 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.compbiolchem.2013.10.005
and Akt3) share >85% homology however, several studies have shown that they are functionally distinct (Lindsley et al., 2005; Breitenlechner et al., 2004). Many cellular processes such as cell survival and insulin signaling are affected by Akt (Green et al., 2008; Garofalo et al., 2003). Various cancer types involving leukemia, breast, pancreatic, ovarian and prostate are caused by activation of Akt due to genetic mutations of its own and/or its interrelated upstream and downstream proteins (PTEN, PI3K) (Saini et al., 2013; Hahne et al., 2013; Zhang et al., 2013; Blake et al., 2012). Thus, a large number of small molecule inhibitors targeting Akt have been developed for suppressing its activation or overexpression (Hers et al., 2011; Lindsley, 2010). Structural studies on Akt isoforms in the apo (Milburn et al., 2003; Yang et al., 2002) or inhibitor bound forms (Bencsik et al., 2010; Kallan et al., 2011) reveal that Akt consist of three conserved domains, an N-terminal pleckstrin homology (PH) domain (residues 1–133), a central kinase catalytic (CAT) domain (149–408) and a C-terminal extension (EXT) domain (residues 449–480) (Calleja et al., 2009b). The interaction of PH and kinase domains keeps Akt1 at inactive conformation in basal conditions (PH-in). Upon stimulation, interaction of PH-domain and phosphoinositides at the plasma membrane triggers conformation change (PH-out)
2
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
resulting in phosphorylation of Thr 308. Phosphorylation of both Thr 308 on the activation loop of kinase domain and Ser 473 on the hydrophobic motif of the C-terminal part results in full activation of Akt1 (Calleja et al., 2009a). Inhibitors against Akt differ in the target sites on the protein (Lindsley, 2010). The inhibitors have been found to bind to ATP binding pocket in the catalytic domain (Blake et al., 2012; FreemanCook et al., 2010; Wang et al., 2011), to the PH domain (Moses et al., 2009; Hartnett et al., 2008), or to the region between PH and kinase domains (Calleja et al., 2009a, 2009b; Wu et al., 2010). To date, most of the inhibitors against Akt have targeted the ATP binding site in the catalytic kinase domain (Garuti et al., 2010). However, due to the high degree of homology (over 50% sequence identity) between the kinase catalytic domains of serine/threonine kinases, selectivity remains a major issue (Rhodes et al., 2008; Luo et al., 2005). Therefore, targeting the PH domain, which shares about 30% identity with other serine/threonine kinases, is an attractive alternative in the development of more selective inhibitors (Barnett et al., 2005; Cherrin et al., 2010). Several studies have been reported focusing on identification of allosteric inhibitors for Akt by computational or experimental techniques (Lindsley et al., 2005; Wu et al., 2008a, 2010; Du-Cuny et al., 2009; Mahadevan et al., 2008). Molecular docking has been used to identify compounds that would selectively bind the PH domain of Akt1 (Du-Cuny et al., 2009; Mahadevan et al., 2008). High throughput screening has also been used by Lindsley et al. (2005), to select allosteric Akt inhibitors that display no inhibition against Akt1 mutants lacking the PH domain, suggesting inhibition of Akt through an allosteric site and resulting in development of Akt1 selective allosteric inhibitor 16 h, which is later called Inhibitor VIII (Lindsley et al., 2005). Calleja et al. (2009b) focused on the catalytic-PH domain interface and used modeling, biochemical assays, FRET to elucidate the mechanism of this allosteric inhibition by compound 16 h of Lindsley’s. Crystallographic structure determination of the complex between Akt and Inhibitor VIII revealed that binding of the inhibitor to the catalytic-PH domain interface ‘locks’ Akt1 in the inactive PH-in conformation by interacting with Trp 80 of the PH domain (Wu et al., 2010). In a recent study of Ashwell et al. (2012), a set of small molecule Akt1 inhibitors which target the same allosteric site with Inhibitor VIII has been determined and optimized by biochemical and biophysical screening strategy followed by crystallization of the complex. In vitro pharmacodynamic and pharmacokinetic studies were followed by identification and crystallization of an allosteric Akt inhibitor that can inhibit intracellular Akt activation (PDB: 4EJN). Here, we build on the recent work on allosteric Akt1 inhibitor development to find novel allosteric Akt1 inhibitors targeting the cavity between kinase and PH domains by combining database filtering with ligand and structure based pharmacophore models and docking. The similarity amongst the hits was considered to obtain a set of compounds. Considering interactions with Akt1, fitness to pharmacophore hypothesis, drug-likeness and ADME properties, derivatives of 3-methyl-xanthine, quinoline-4-carboxamide and 2-[4-(cyclohexa-1,3-dien-1-yl)-1H-pyrazol-3-yl]phenol were proposed as lead scaffolds. This study is the first time application of virtual screening methodologies for identification of allosteric inhibitors targeting the cavity between kinase and PH domains of Akt1.
2. Methods Schrödinger’s Small Molecule Drug Discovery Suite (2011) was used with Linux operating system on a HP xw6600 Workstation for all virtual screening applications. Maestro was the unified graphical interface for access to all Schrödinger modules; LigPrep (LigPrep,
2012), Protein Preparation Wizard (PPW, 2012; Madhavi Sastry et al., 2013), Prime (Prime, 2012a), ConfGen (Watts et al., 2010; ConfGen, 2012), Phase (Phase, 2012b; Dixon et al., 2006a, 2006b), QikProp (QikProp, 2012) and Glide (Glide, 2012; Friesner et al., 2004; Loving et al., 2009). 2.1. Receptor protein structure preparation The structure file from the Protein Data Bank (PDB) was prepared for use in the molecular modeling calculations. In the PDB, there was only one Akt1 structure with the ligand bound to the allosteric region between kinase and PH domains at the time this study initiated (PDB code: 3O96). Protein Preparation Wizard (PPW) was used to assign bond orders, add missing hydrogens to heavy atoms and predict coordinates of missing loops (residues: 45–49, 88–92, 188–199, 298–313) (Prime, 2012a). Only water 455 was kept due to its bridging function between primary amine and Asn54 (Kettle et al., 2012) and all other water molecules were deleted. Five protonation states were generated for the native ligand (pH 7.0 ± 4.0) (Schrödinger, 2012) and the state with the lowest state penalty of 0.18 kcal/mol and charge of +1 was selected. Hydrogen-bonding network was optimized by reorienting hydroxyl and thiol groups, amide groups of asparagine and glutamine and the imidazole ring in histidine. Protonation states of histidine, aspartic acid and glutamic acid, and tautomeric states of histidine were predicted at neutral pH. Finally, the structure was refined by minimization (Impref utility of Impact module) using OPLS 2005 force field with a maximum RMSD of 0.30 A˚ for heavy atoms. 2.2. Binding site selection Receptor binding site was defined by grid generation using Glide ˚ and (Glide, 2012). Each field on the grid is identical in size (1 A) represents the shape and properties of the active site. Glide uses two cubic boxes to define where the energetic calculations should be done. The outer cube contains all atoms of the binding ligand and all the acceptable positions of the ligand center lies in the inner cube. The centers of both the outer cube of 20 A˚ size and the inner cube of 10 A˚ size were chosen as the centroid of the native ligand, Inhibitor VIII (Wu et al., 2010) (PDB code: IQO) of the complex (PDB code: 3O96). Van der Waals radii of non polar atoms were scaled to 0.8 with 0.25 partial charge cutoff to provide flexibility to the rigid receptor and to relax binding site, which is a buried cavity between PH and kinase domains of the protein, by decreasing penalties for close contacts in the binding site. 2.3. Generation of 3D database Drug-like compounds of 38 big vendors in ZINC database (Irwin et al., 2012) were obtained (Table S1). Ionization and tautomer states of molecules were generated using Epik module at a pH range of 7 ± 2. Stereochemical information was obtained from the 3D geometry of the structures. Starting chiralities were selected by LigPrep (2012), based on the chemistry of naturally occurring steroids, fused rings and peptides. Up to 32 stereoisomers were generated and filtered to retain those with the lowest energies. Conformers were generated by ConfGen (ConfGen, 2012) module with a maximum conformer number of 100 per compound and 10 per rotatable bond. Energy minimization was carried out in 50 steps for the conformers using the OPLS 2005 force field and a distance dependent dielectric constant of 4. Molecular, drug-likeness and ADME (absorption, distribution, metabolism, excretion) properties of compounds were predicted by QikProp. Only the structures satisfying Lipinski’s Rule of Five (Lipinski et al., 2001) were kept. Structures that had reactive groups were removed by ligfilter script.
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
Database generation resulted in 499,026 compounds with ∼26 million conformations. 2.4. Structure-based pharmacophore modeling E-pharmacophores script (Loving et al., 2009; Salam et al., 2009) was used for pharmacophore modeling based on the allosteric binding site of Akt1 receptor. Schrödinger fragment library (667 small fragments with accessible low energy ionization and tautomeric states from 441 unique small fragments; 1–7 ionization/tautomer variants; 6–37 atoms; MW range 32–226) was docked to grid generated Akt1 structure using Glide XP mode. E-pharm script used the information stored in Glide XP descriptor file, which carries the information about fragment binding modes to receptor and energetic terms, to sum the energetic contributions that construct GlideScore of all fragments on each atom center. The summed values allowed detection of seven possible pharmacophore sites; R2246, H1462, A575, H1384, D937, D1305 and A598 (hydrogen bond acceptor (A), hydrogen bond donor (D), hydrophobic group (H), negatively charged group (N), positively charged group (P), and aromatic ring (R)) on atoms. These sites were ranked based on their energies and two pharmacophore hypotheses were developed with selected sites of R2246, H1462, H1384 (RHH hypothesis) and of R2246, A575, D937 (RAD hypothesis). R2246 was included in both hypotheses to satisfy the pi-stacking interaction between the ligand and Trp80 of Akt1. RHH would select for ligands that would have similar hydrophobic character with the Inhibitor VIII triple ring substructure, while RAD hypothesis promised a different chemical scaffold from Inhibitor VIII with D937 4.1 A˚ away from it. Generated ZINC database was filtered separately to 10,000 compounds using RHH and RAD hypotheses for reducing the number of structures to be used in two separate docking runs.
3
of vector features (acceptors, donors, and aromatic rings) in the aligned structures. Site score is the root-mean-squared deviation in the site-point positions. Volume score is based on the overlap of van der Waals spheres of the non-hydrogen atoms in each pair. After scoring, the APRRR hypothesis with the third highest survival score was selected since the first two hypotheses exhibited low fitness to native ligand (Inhibitor VIII). 2.6. Docking and scoring The virtual docking experiments were performed using Glide module with two docking modes; standard precision (SP) and extra precision (XP). For scoring the binding affinity of the poses, Glide employs its own scoring function GlideScore (GScore). Two different docking runs were performed for each hypothesis separately. Docking 10,000 ZINC compounds obtained by filtering with RHH and RAD hypotheses to grid generated Akt1 structure with SP mode and scoring with SP scoring function was followed by re-docking top scored 1000 molecules (10%) using XP docking mode and scoring with XP scoring function. The numbers 10,000 and 1000 were determined as optimum values considering the CPU time required for docking calculations. For both docking modes, ten poses per compound state (at maximum) were generated and only best scoring state was kept in the final hit list. Strain-Rescore script was applied to identify ligands with too much strain. For each ligand pose, a tightly constrained minimization and an unconstrained minimization were performed with MacroModel module. Ligands with more than 5 kcal/mol energy difference between the docked and free conformations received penalties and quarter of this energy difference was added to the GlideScore. Strain-corrected GlideScore values were used for further evaluation of the hits. 2.7. Similarity analysis
2.5. Ligand based pharmacophore modeling To construct a ligand based pharmacophore hypothesis, six small molecule allosteric Akt1 inhibitors, obtained from several studies in literature (Lindsley et al., 2005; Hartnett et al., 2008; Du-Cuny et al., 2009; Mahadevan et al., 2008; Wu et al., 2008a) and Inhibitor VIII (Wu et al., 2010) were used. The information of these ligands was used to identify a set of pharmacophore site points (features) and model hypotheses by Develop Pharmacophore Model workflow of Phase module (Phase, 2012b). 2D structures of the ligands were drawn by ChemSketch software (ACD, 2013), followed by conversion to low-energy 3D structures with LigPrep module (LigPrep, 2012), where hydrogen atoms were added, protonation was done to produce ionization states at neutral pH, maximum 32 stereoisomers were generated for each ligand, and an energy minimization was performed to generate maximum 100 conformers for each ligand by ConfGen (2012). Twelve possible combinations of the seven pharmacophore features (A, D, H, N, P, R), namely ‘variants’ were detected by Phase. The number of the pharmacophore features in a variant can be minimum three and maximum seven. In this study, minimum four and maximum seven site points were selected to be included in a variant and all seven ligands were set to be match each site point. These twelve variants allowed detection of 413 hypotheses with seven common pharmacophore models including five pharmacophore sites (AADRR, PRRRR, APRRR, DRRRR, AAADR, AAPRR, ADRRR). The number of hypotheses contributed to each common pharmacophore model varied between 1 and 129. The model which had the highest number of hypotheses contribution, APRRR, was scored to find optimum alignments to active set of inhibitors. Assessment of the models was based on survival score; which is a combination of (default weighting factor of 1.0) vector score, site score and volume score. Vector score is the average cosine of the angles formed by corresponding pairs
Similarity between docking hits was conducted to identify a diverse set of chemical scaffolds as potential inhibitors using ChemMine Tools (Backman et al., 2011). The similarity measure was based on the Tanimoto Coefficient (TC). The Tanimoto coefficient has a range from 0 to 1 with higher values indicating greater similarity than lower ones. SMILES codes of ligands were provided to tool and resulting distance matrix was passed to the hierarchical clustering program automatically. The program hierarchically joins the most to least similar compounds using single linkage. A tree file was generated to visualize two-dimensional similarity of ligands and identify clusters of common scaffolds. Five clusters were detected to be analyzed in detail for structure-based models; RADfamily1 to RAD-family4 and RHH-family5 and clustering was not applied to hits of ligand-based pharmacophore model due to small number of surviving compounds. 2.8. Prediction of ADME and druglikeness properties ADME (absorption, distribution, metabolism, excretion) properties and Lipinski’s Rule of Five were predicted by QikProp (QikProp, 2012). QikProp provides acceptable ranges for comparing the predicted properties of compounds with those of 95% of known drugs. The investigated ADME properties were PSA (van der Waals surface area of polar nitrogen and oxygen atoms), percentage of human oral absorption (HOA%), SASA (total solvent accessible surface area in Å2 ), FOSA (hydrophobic component of the SASA, associated with a saturated carbon and attached hydrogen), FISA (hydrophilic component of the SASA), PISA ( component of the SASA, is referred to a carbon and attached hydrogen), WPSA (weakly polar component of the SASA, takes the polar components as electronegative atoms such as nitrogen and oxygen; and halogens, phosphorus and sulphur atoms). Lipinski’s Rule of Five (Lipinski et al., 2001) was used to
4
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
evaluate drug-likeness or determine if a chemical compound with a certain pharmacological or biological activity has properties that would make it a likely orally active drug in humans. 2.9. Induced Fit Docking Induced Fit Docking (IFD) protocol, provided by Schrödinger was used to introduce side chain flexibility to receptor Akt1 during ligand docking. The multistep protocol starts with Glide XP docking to rigid receptor with a vdW radii scaling of 0.5 for both the ligand and the receptor. Next, receptor side chain flexibility was provided using Prime module by generation of up to 20 poses for each ligand and residues which had at least an atom at a distance of 5 A˚ from any of these 20 poses were subjected to undergo conformational change while coordinates of other residues remain fixed. In the final step, the ligands were redocked to low energy induced fit structures in Glide XP mode with a vdW scaling of 1.0 (no scaling) and receive their final scores according to XP scoring function. 2.10. PH domain sequence alignment The BLAST algorithm (Altschul et al., 1990) was used to align PH domain sequence of Akt1 to PH domain sequences of seven kinases (Table S9) in order to find the similar sequences and identify the conservation of important interacting residues. Standard protein–protein blast (blastp) was used with query sub-range of (1–133) for Akt1 and default algorithm parameters of blastp. Fig. 1. Workflow of the methodology.
3. Results The potential allosteric inhibitors that were expected to bind the PH-kinase domain cavity of Akt1 were identified by using pharmacophore modeling and docking approaches. Pharmacophore models were built either based on the information of the binding site of the receptor or based on the previously identified allosteric inhibitors for Akt1. ZINC drug-like database was filtered based on 3D similarity to the pharmacophore hypotheses. Two-step docking was applied using Glide to predict binding modes and affinities of compounds in the filtered database. Common scaffolds in the top scoring hits were identified using clustering based on the Tanimoto similarity coefficient (Backman et al., 2011). The compounds from the selected clusters were further filtered based on interaction with Trp 80, fitness to pharmacophore hypothesis, ADME, druglikeness and strain penalty criteria. The surviving hits were analyzed in detail to propose a set of novel allosteric inhibitors for Akt1. The workflow of the study is shown in Fig. 1. 3.1. Preparation of Akt1 allosteric binding site Following addition of coordinates of hydrogen atoms and assignment of bond orders, missing loops of Akt1 (PDB code: 3O96) were predicted (residues: 45–49, 88–92, 188–199, 298–313). The coordinates for the Gal 113-Val 145 loop could not be predicted since Prime is limited to predict loops with fewer than 20 residues. However, this loop was not in the vicinity of the binding site, therefore its presence was not considered to be crucial for docking calculations. The loop between Ala 188 and Asn 199, which includes ␣C-helix of Akt1, was predicted by Prime. The correct positioning of ␣C-helix and DFG motif is necessary for allosteric inhibition by locking Akt1 in inactive PH-in configuration (Palmieri and Rastelli, 2013). In the allosteric inhibitor bound structure, only the coordinates for the N-terminal residues of the ␣C-helix are resolved. Compared with other kinase structures from different families, the location of these residues suggests that the helix assumes a “turnedout” conformation which is more than 7 A˚ away from Inhibitor VIII
and the allosteric binding pocket. Because the ligand–residue inter˚ the predicted action range of energy function (GlideScore) is 3 A, coordinates of the ␣C-helix were not a part of the docking calculations. Only water 455 was maintained out of 21 water molecules for further calculations due to its bridging function between ligand and Asn54 of Akt1 (Kettle et al., 2012). Hydrogen bonding network was optimized and minimization was done on the heavy atoms by using OPLS 2005 force field with a maximum allowed RMSD of 0.3. Taking Inhibitor VIII (native ligand) as reference, the binding ˚ site was selected with two co-centered cubic boxes (inner: 10 A, ˚ The van der Waals scaling value of nonpolar atoms was outer: 20 A). taken as 0.8 with 0.25 partial charge cutoff to mimic the presence of a larger binding site allowing larger ligands to be docked. Fig. 2 shows Akt1 with Inhibitor VIII and close-up view of the binding site. 3.2. Docking of previously identified allosteric inhibitors 2D structures of the previously identified eight allosteric Akt inhibitors (Hartnett et al., 2008; Wu et al., 2008a, 2010; Du-Cuny et al., 2009; Mahadevan et al., 2008) and Inhibitor VIII were transformed into 3D by LigPrep (2012) with varying ionization and tautomerization states at neutral pH. For nine inhibitors, maximum 25 stereoisomers were created and energetically minimized conformers were generated by ConfGen (2012). Conformers whose energy were more than 25 kcal/mol higher than the lowest energy conformer were eliminated. Finally, 391 conformers were obtained for nine inhibitors (including Inhibitor VIII) and were docked to allosteric binding site of Akt1 using Glide XP without applying any constraints. For each inhibitor, the conformer which showed the highest binding affinity (based on GScore value) for Akt1 was examined in detail (Fig. S1). Glide XP GScores for the highest ranked conformer of each ligand, its rank among all docked conformers of nine ligands and the key interactions they make with Akt1 are shown in Table 1. The docking results showed that Inhibitor VIII had the best docking score of −10.94 kcal/mol, which was a validation of
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
5
Table 1 Experimentally identified allosteric AKT1 inhibitors. Assigned name
Chemical structure
AKT1 IC50
AKT2 IC50
Glide XP GScore (kcal/mol)
Interactions with AKT1
INHIBITOR VIII (Arencibia et al., 2013)
58 nM
210 nM
−10.94
Trp 80, Arg 273, Water 455
Compound1 (Calleja et al., 2009a)
760 nM
24000 nM
−10.0
Trp 80, Arg 273, Water 455
Compound2 (Calleja et al., 2009a)
212 nM
325 nM
Compound3 (Hennessy et al., 2005)
126 nM
Compound4 (Liu et al., 2009)
−9.95
Trp 80, Arg 273, Water 455
22 nM
−10.14
Arg 273, Water 455, Glu 85
18 nM
239 nM
−10.10
Trp 80
Compound5 (Liu et al., 2009)
21 nM
85 nM
−10.27
Trp 80, Arg 273, Water 455, Tyr 326
Compound6 (Liu et al., 2009)
14 nM
99 nM
−10.79
Trp 80, Arg 273, Water 455, Glu 85
Compound7 (Lindsley et al., 2005)
13 M
–
−9.91
Trp 80, Asn 53
Compound8 (Breitenlechner et al., 2004)
6.3 M
–
−7.75
Thr 211
6
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
Fig. 2. (A) Akt1 structure in complex with Inhibitor VIII (red) bound to the allosteric binding site at the interface of the kinase (blue) and PH (green) domains (PDB code: 3O96). (B) Close-up view of the allosteric binding site. Inhibitor VIII (red), Trp 80 (orange), Water455 (black) and other important residues are shown as sticks in varying colors. Same color scheme is used in the other structure figures. All figures are prepared in Maestro interface. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)
the docking protocol. The remaining ligands exhibited good binding affinity to the allosteric site except Compound8. Compound8 has a bulky dodecane tail which inhibited its fitting in the cavity between kinase and PH domains of the protein. Six of the eight ligands made the – stacking interaction with Trp 80 residue of Akt1 necessary for the allosteric mechanism of inhibition (Wu et al., 2010). Investigation of structural similarity of the ligands indicated that Compound1 and Compound2 shared a common {6-phenyl5-[4-(piperidin-1-ylmethyl)phenyl]-1,2-dihydro pyrazine} substructure while Compounds 3–6 shared a common {3-phenyl-2[4-(piperidin-1-ylmethyl)phenyl] pyridine sub-structure. Inhibitor VIII has a {2-phenyl-3-[4-(piperidin-1-ylmethyl)phenyl] pyrazine} sub-structure, whereas Compounds 7 and 8 showed no structural similarity to other 6 ligands and Inhibitor VIII. Based on the interaction diagrams of all inhibitors, Inhibitor VIII, Compound1 and Compound2 were observed to make the – stacking interaction with Trp 80 through their pyrazine ring. The same interaction was observed between pyridine ring and Trp 80 for Compounds 4, 5 and 6. The -cation interaction with Arg 273 was observed between (1,3-dihydro-2H-benzimidazol-2-one) sub-structure of Compounds 1, 2 and Inhibitor VIII. Although there was no information on binding modes and exact binding sites for these eight compounds, six of them showed good binding affinity in virtual docking and made favorable interactions within the allosteric binding site. 3.3. Identification of potential inhibitors by structure-based pharmacophore modeling 3.3.1. Structure-based pharmacophore modeling Structure-based pharmacophore sites were identified using unconstrained fragment docking to the allosteric binding site of Akt1. The descriptive information from energetic terms and binding modes of 667 conformers of 441 unique fragments was used to determine the seven pharmacophore sites (site score ≥2.0) by Epharmacophores script of Phase module. Ranks, types, scores and coordinates of seven sites are given in Table S2. Hypotheses composed of three sites were selected to allow sampling of a wider range of compounds because as the number of sites increases, the pharmacophore model becomes more restrictive. Sites with higher energy contribution were included if they were less than 10 A˚ away from each other. The only ring feature detected among the seven sites, R2246 was included to assure the -stacking interaction between the ligand and Trp 80 of Akt1 (Wu et al., 2010). Consequently, one hypothesis was created using R2246, H1462, H1384 (RHH hypothesis) and another one was created using R2246,
A575, D937 (RAD hypothesis). Locations of seven detected pharmacophore sites within the binding site and inter-site distances for RHH and RAD hypotheses are shown in Fig. 3. Note that RHH would select for ligands that would have similar hydrophobic character with the triple ring substructure of Inhibitor VIII, while RAD hypothesis would select ligands with a different chemical scaffold from Inhibitor VIII. Pharmacophore filtering with RHH and RAD hypotheses were performed separately for the generated ZINC database and 10,000 compounds with the highest fitness scores were saved for each hypothesis. 3.3.2. Docking of compounds selected with pharmacophore models RHH and RAD to full length Akt1 10,000 compounds obtained separately by filtering the ZINC database with structure-based RHH and RAD hypotheses were initially docked to Akt1 allosteric site in SP mode and scored with SP scoring function. The 10% top scoring (∼1000) molecules were redocked using XP docking. This two step unconstrained docking of the initial 10,000 compounds resulted in 997 docked ligands for the RHH hypothesis and 1000 docked ligands for the RAD hypothesis. The hits obtained by RAD hypothesis had better GScore values (−12.2 to −8.3 kcal/mol) compared to the hits obtained by RHH hypothesis (−11.4 to −2.3 kcal/mol). 997 docked ligands of RHH hypothesis showed a sharp decrease in their GScore values from −11.4 kcal/mol to almost −9 kcal/mol and after −9 kcal/mol, the decrease in GScore values was significantly smaller. Therefore, cutoff GScore value was chosen as −9.00 kcal/mol for docking hits of RHH hypothesis. In order to be consistent, the same cutoff value of −9.00 kcal/mol was also applied for docking hits of RAD hypothesis. 88 hits were obtained from RHH hypothesis and 630 hits were obtained from RAD hypothesis. When the strain energy values were considered, 38 compounds based on the RHH hypothesis and 158 compounds based on the RAD hypothesis survived and were further analyzed. 3.3.3. Post-docking analyses of compounds in RAD and RHH clusters The selected hits (total 196) (Section 3.3.2) were first evaluated based on chemical similarity in order to determine the types of scaffolds selected by pharmacophore filtering and docking. To this end, ChemMine Tools (Backman et al., 2011) were used to identify common scaffolds between the docked ligands to Akt1 structure with single-linkage hierarchical clustering based on Tanimoto coefficient (Ajay et al., 1998). The highly populated clusters indicate enrichment of these similar compounds in the top scoring compounds after docking. The similarity tree including 196 compounds
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
7
Fig. 3. (A) Location of seven detected pharmacophore sites in the Akt1 allosteric binding site. (B) Inter-site distances for RHH hypothesis (left) and RAD hypothesis (right). Ring in orange, hydrophobic sites in green, hydrogen bond acceptor sites in red and hydrogen bond donor sites in blue. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)
for hits of RHH and RAD hypotheses revealed five significant clusters; four clusters for RAD hits (RAD-1; 27 hits, RAD-2; 20 hits, RAD-3; 12 hits, RAD-4; 22 hits) and one cluster for RHH hits (RHH; 13 hits). The interaction diagrams of all compounds in clusters were obtained and key protein–ligand interactions were identified. The – stacking interaction with Trp 80 of Akt1 was checked for members of all clusters. This interaction was previously suggested to be compulsory for allosteric mechanism of inhibition (Wu et al., 2010). Compounds in RAD-2, RAD-4 and RHH clusters interact with Trp 80, whereas compounds in RAD-1 and RAD-3 clusters do not and were eliminated from further analysis. For all compounds in clusters RAD-2, RAD-4 and RHH, no violations to Lipinski’s Rule were observed. Eight compounds in RAD-2 cluster, two compounds in RAD-4 cluster and two compounds in RHH cluster indicated low values for WPSA. Seven compounds in RAD-4 cluster indicated PISA values higher than 95% of known drugs. All compounds exhibited acceptable human oral absorption values. The compounds that were out of ADME and druglikeness ranges were eliminated. As a result, nine compounds in RAD-2 cluster, twelve compounds in RAD-4 cluster and five compounds in RHH cluster survived. The descriptor values were given in corresponding property tables with their ranges for surviving compounds in RAD-2, RAD-4 and RHH in supplementary file (Tables S3–S5). The ‘fitness score’ measures the proximity of each compound to matching pharmacophore site points and how well the matching vector features such as acceptors, donors and aromatic rings overlay with the hypothesis and it ranges from −1.0 to 3.0. The fitness scores of compounds varied between 1.93 and 2.02 for RAD-2 cluster, 1.94 and 2.22 for RAD-4 cluster and 1.11 and 1.36 for RHH cluster (Table S6). Since the number of pharmacophore sites in a hypothesis was kept at minimum, the ligands were expected to align well with the pharmacophore model. All compounds in RHH cluster had significantly low fitness scores than those in the RAD clusters, which displayed better alignment through visual analysis. Consequently, nine compounds (all) in RAD-2 cluster and twelve compounds (all) in RAD-4 cluster were preserved while RHH cluster was discarded. Prior to proposing final hits, compounds that had better than −9.0 kcal/mol GScores were investigated to check if any strain penalties were received and such compounds were eliminated. Ligands with more than 5 kcal/mol energy difference between the docked and free forms received penalties and quarter of the strain energy difference was added to the GScore. Finally, seven compounds survived from RAD-2 cluster and eleven compounds survived from RAD-4 cluster. GlideScores and strain corrected GlideScores for RAD-2 and RAD-4 clusters were shown in Table S7.
3.4. Identification of potential inhibitors by ligand-based pharmacophore modeling 3.4.1. Ligand-based pharmacophore modeling Previously identified small molecule allosteric inhibitors (Hartnett et al., 2008; Du-Cuny et al., 2009; Mahadevan et al., 2008; Wu et al., 2008b) were used to determine ligand based pharmacophore models. Compounds 1–6 (Table 1) and Inhibitor VIII provided detection of twelve possible combinations of pharmacophore features. Pharmacophore model APRRR was selected such that each pharmacophore site matched all seven ligands, the model exhibited a high fitness to Inhibitor VIII and the ring structure that interacts with Trp 80 on Akt1 was included in the model (Fig. 4). The model exhibited full fitness to Compound6 (score: 3.00/3.00) and the second best fitting compound was Inhibitor VIII (score: 2.71/3.00). The tricyclic feature set forms a hydrophobic moiety in the allosteric binding site, which holds for most of the type II kinase inhibitors (Garuti et al., 2010). Feature R14 makes – stacking interaction with Trp 80 and feature P12 stands for the primary amine (NH+ ) interacting with Water 455. Filtering the database with APRRR hypothesis based on 3D similarity yielded 701 matching compounds with positive fitness scores.
Fig. 4. Alignment of ligand-based pharmacophore hypothesis APRRR on best fitting compound 6.
8
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
The 701 ligands, that matched the ligand-based APRRR hypothesis, were docked to Akt1 allosteric site using XP mode. GScore values of 701 compounds varied between −10.7 to −0.5 kcal/mol. The same GScore cutoff value of the hits of RAD and RHH models, −9.00 kcal/mol was also applied for docking the hits of APRRR hypothesis resulting in 64 hits. 16 compounds based on the APRRR hypothesis survived after strain energy correction and were further analyzed.
fitness scores of the final two hits of APRRR hypothesis are given in Table S6.
3.5. Interaction analyses of proposed compounds Eliminations considering protein–ligand interactions, ADME and druglikeness criteria, fitness to pharmacophore hypothesis and strain energy penalties resulted in seven final compounds in RAD-2 cluster, eleven compounds in RAD-4 cluster and two compounds of APRRR hypothesis. Compound names, ZINC Codes, GScore, fitness score and ADME property values for final 20 hits of 3 pharmacophore models are listed in Table 2. The IUPAC names for the common sub-structures of compounds in RAD-2, RAD-4 and APRRR hits were determined by using ChemSketch Nomenclature software (ACD, 2013). Seven surviving compounds in RAD-2 cluster all shared a 3methyl-xanthine scaffold (Fig. S2). A representative compound with the highest GScore and fitness values, and its interactions with the allosteric site are shown in Fig. 5a. 1, 2, 4, 5, 6 and 7 in RAD-2 cluster make – stacking interactions with Trp 80 through their imidazole ring of xanthine sub-structure and hydrogen bonding interactions between Ser 205 and primary nitrogen (NH) of xanthine. 3 shows – stacking interaction with Tyr 272 through its imidazole ring of xanthine sub-structure and interacts with Gln 79 through a hydrogen bond. 1, 2 and 6 make H-bonds with Asp 292 side chain. The final eleven compounds in RAD-4 cluster shared a quinoline4-carboxamide scaffold (Fig. S3). A representative compound with the highest GScore and fitness values, and its interactions with the allosteric site are shown in Fig. 5b. All members of RAD-4 cluster make – stacking interaction with Trp 80 through their quinoline sub-structure. Except for 9 and 18, all compounds also make hydrogen bonding interaction with side chain of Gln 79 and the conserved
3.4.2. Post-docking analyses of APRRR hits The interaction diagrams of the 16 hits obtained from APRRR pharmacophore model were acquired and key protein–ligand interactions were identified. The – stacking interaction with Trp 80, which is compulsory for allosteric mechanism of inhibition of Akt1 (Garuti et al., 2010) was checked. 11 of 16 hits of APRRR hypothesis were observed to make the – stacking interaction with Trp 80 and were kept for further analysis. All compounds obeyed to Lipinski’s Rule. 6 of 11 compounds indicate low values for WPSA compared to 95% of known drugs. All compounds exhibited acceptable human oral absorption values. The compounds that were out of ADME and druglikeness ranges were eliminated resulting in survival of five hits of APRRR hypothesis. The values and ranges of ADME descriptors were given in corresponding property tables for surviving compounds (Table S8). Three pharmacophore sites were included in structure-based pharmacophore models RHH and RAD and since matching three sites is easier than matching five pharmacophore sites with the compounds, the compounds filtered with structure based models exhibited significantly higher fitness scores than compounds filtered with ligand based pharmacophore model APRRR. Therefore, threshold value was defined as 1.0 for APRRR hypothesis. Considering this threshold value, two compounds were preserved. The
Table 2 Compound names, ZINC codes, GScore, fitness score and ADME property values for final 20 hits of 3 pharmacophore models. Detailed descriptions of the compound properties are given below the table. Compound name RAD-2
ZINC code
GScore (kcal/mol)
Fitness score
SASA (Å2 )
FOSA (Å2 )
FISA (Å2 )
PISA (Å2 )
WPSA (Å2 )
PSA (Å2 )
HOA%
1 2 3 4 5 6 7
ZINC00839381 ZINC02373214 ZINC13512705 ZINC00625908 ZINC00850903 ZINC00839391 ZINC00633233
−10.21 −10.11 −9.73 −9.43 −9.16 −9.02 −9.01
1.98 1.98 1.96 2.02 1.99 1.97 1.93
697.35 691.11 802.37 745.23 678.66 698.02 682.65
267.91 221.87 196.21 358.77 356.37 269.26 343.97
183.85 196.52 157.09 134.55 140.42 182.61 137.71
227.56 179.77 377.37 180.25 170.33 227.98 179.55
18.03 92.94 71.69 71.67 11.54 18.16 21.42
133.60 135.22 132.04 121.97 109.95 132.95 111.05
79.04 77.26 93.06 96.88 93.10 79.32 92.46
RAD-4 8 9 10 11 12 13 14 15 16 17 18
ZINC16392469 ZINC16392736 ZINC13321319 ZINC02753923 ZINC01073862 ZINC02867568 ZINC08917432 ZINC01073706 ZINC00656127 ZINC13321889 ZINC01068312
−10.17 −9.82 −9.76 −9.63 −9.63 −9.59 −9.51 −9.47 −9.45 −9.28 −9.19
ZINC09410511 ZINC09410451
−10.08 −9.78
2.22 2.06 1.94 2.10 2.12 1.95 1.95 1.95 2.23 2.01 2.11
753.98 726.33 720.11 745.39 748.59 705.96 712.30 761.06 712.19 735.51 744.46
334.14 183.82 179.14 278.51 311.65 150.31 150.16 267.57 220.49 207.46 146.84
113.05 190.74 129.83 107.73 148.72 126.87 119.03 135.55 126.12 136.94 179.28
265.99 349.85 369.89 318.99 256.33 390.28 351.90 247.66 333.86 350.89 383.61
40.80 1.92 41.24 40.17 31.90 38.50 91.21 110.28 31.72 40.22 34.72
108.22 128.79 99.46 107.84 103.05 90.86 90.65 102.58 110.66 101.85 123.91
100.00 81.32 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 90.83
APRRR-1 19 20
1.33 1.80
755.46 781.24
221.15 302.69
201.99 199.22
260.63 207.65
71.69 71.69
113.66 113.12
61.32 63.98
SASA: Total solvent accessible surface area, FOSA: Hydrophobic component of the SASA; associated with a saturated carbon and attached hydrogen), FISA: Hydrophilic component of the SASA; associated with the solvent accessible surface area on N, O, and H on heteroatoms, PISA: component of the SASA; is referred to a carbon and attached hydrogen, WPSA: Weakly polar component of the SASA; takes the components as electronegative atoms such as N, O and halogens, P and S atoms, PSA: Van der Waals surface area of polar N and O atoms, HOA%: Percentage of human oral absorption. Fitness Score (unitless): Measure of proximity of each compound to matching pharmacophore site points (vector features such as acceptors, donors and aromatic rings); ranges from −1.0 to 3.0.
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
9
Fig. 5. Binding modes and important binding site residues of (A) 1 (RAD-2 cluster), (B) 8 (RAD-4 cluster), (C) 20 (APRRR hit) located in allosteric binding site of Akt1 (ligands in magenta, PH domain in green, kinase domain in blue), (D) superimposition of 1, 8 and 20 on Inhibitor VIII (red). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)
Table 3 Interactions with Akt1 for 20 final hits in RAD-2, RAD-4 and hits of APRRR model. RAD-2 Cluster
1 2 3 4 5 6 7
Trp 80
Ser 205
Asp 292
pi pi pi pi pi pi pi
H (b) H (b)
H (s) H (s)
Tyr 272
Thr 81
Lys 268
H (b)
H (b)
H (b) pi
H (b) H (b) H (b) H (b)
Gln 79
H (s)
RAD-4 cluster
8 9 10 11 12 13 14 15 16 17 18
Trp 80
Lys 268
pi pi pi pi, pi pi pi pi pi pi pi pi
H (s)
Asp 292
Tyr 272
Gln 79
Water 455
H (s)
H (s)
H (s) H (s) H (s) H (s) H (s) H (s) H (s) H (s)
H (s) H (s) H (s) H (s) H (s) H (s) H (s) H (s)
H (s)
H (s) pi pi
H (s)
Thr 82
pi H (b)
APRRR hits
19 20
Trp 80
Tyr 272
Thr 211
pi, pi pi
pi pi
H (b) H (b)
Ala 58 H (b)
Leu 78 H (b)
Asn 53
Ser 205
H (b)
H (b)
10
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
Fig. 6. (A) Superimposition of IFD poses 1, 2, 3, 4 for compound 1 on rigid docked form (turquoise) and Inhibitor VIII (purple) and IFD conformations of important residues in compound 1 bound forms. (B) Superimposition of IFD poses 1, 2, 3, 4, 7, 8 for compound 8 on rigid docked form (turquoise) and Inhibitor VIII (purple) and IFD conformations of important residues in Compound 8 bound forms. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)
water 455. 10, 11, 17 make – stacking interaction and 18 make hydrogen bond with Tyr 272. 8 and 11 showed hydrogen bonding with side chain of Lys 268. Two hits of APRRR hypothesis shared a 2-[4-(cyclohexa1,3-dien-1-yl)-1H-pyrazol-3-yl]phenol scaffold and show – stacking interaction with both Trp 80 and Tyr 272. A representative compound with the higher GScore and fitness values, and its interactions with the allosteric site are shown in Fig. 5c. 20 makes hydrogen bonds with backbone of Ala 58, backbone of Leu 78 and backbone of Thr 211. 19 makes three hydrogen bonds with backbones of Asn 53, Ser 205 and Thr 211. A summary of the interactions of final 20 compounds is listed in Table 3. Since the allosteric binding site was buried between kinase and PH domains, locating ligands without strain was hard and using information of the binding site with structure-based pharmacophore modeling was more successful on finding ligands which show high affinity to the allosteric binding site. Between the two structure-based models, RAD hypothesis predicted compounds with better binding scores and better alignment with the hypothesis. 3.6. Induced fit docking Induced fit docking was carried out for the molecules with the highest fitness scores and GlideScores from RAD-2, RAD-4 and APRRR hypotheses to see how well the binding conformations and interactions of each hit obtained from docking to rigid protein structure can be reproduced when side chain flexibility is allowed in addition to ligand flexibility during docking. For each ligand, up to 18 ligand conformations were obtained. In addition to investigation of the docking scores and protein–ligand interactions, the IFD models that make -stacking interaction with Trp 80 for each ligand were superimposed to the rigid docked forms to detect the difference in binding modes. The results showed that the difference between GlideScores obtained from IFD and rigid docking did not exceed (±) 1.8 kcal/mol for all ligand–protein complexes. For eighteen IFD complexes of
1 (RAD-2 cluster), the – stacking interaction was observed for five ligands through their imidazole ring of xanthine sub-structure and four of these five ligands also interacted with Ser 205 through a hydrogen bond similar to the rigid docked form. One of these five ligand conformations generated by IFD (IFD4, GlideScore: −9.90 kcal/mol) showed good alignment with rigid docked form of 1 with a RMSD of 3.5 A˚ for all heavy atoms of the ligand. The – stacking interaction was observed for eight ligands out of seventeen IFD complexes obtained for 8 (RAD-4 cluster) through their quinoline sub-structure. Ligands in seven of these complexes made hydrogen bond interaction with Gln 79 and Wat 455 as in the rigid docked form. Three compounds exhibited – stacking interaction with Tyr 272 and four compounds interacted with Lys 268 through a hydrogen bond. The best alignment to rigid docked form was observed for ligand conformation of IFD8 complex with a RMSD of 5.94 A˚ for all heavy atoms of 8. Out of six IFD complexes obtained for 20, three interacted with Trp 80. Overall, binding modes of 1 and 8 were reproduced with small deviations by IFD and similar key protein–ligand interactions were observed; while binding mode of 20 exhibited a larger deviation from rigid docked form. For 1 and 8, IFD poses in Table S9 were further investigated considering conformational closeness and movement of residues in the binding site provided by side chain flexibility of IFD. Superimposition of rigid docked form and five IFD poses of 1 on Inhibitor VIII revealed that IFD 1, 2 and 3 showed good alignment except for the benzene ring of IFD3 at the tail (Fig. 6a). The Trp80 interacting bicyclic parts of the first three IFD poses were also located at close coordinates to Trp80 interacting tricyclic part of Inhibitor VIII. Conformations of rigid docked form and IFD4 of 1 were completely different from IFD1, 2 and 3. IFD5 had completely different binding mode than other IFD models. Examination of residues allowed detection of movements for five binding site residues. The most significant movement was observed for side chain of Trp80 (1H-indene sub-structure). Other re-oriented residues were Ser205, Lys268, Asp292 and Arg273, for which overlapping was observed for IFD1, 2 and 3 except for small deviations in coordinates of Ser205.
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
11
Fig. 7. Binding modes of compounds (A) 1, (B) 8, (C) 20 in allosteric binding site of AKT1; in the absence of PH domain (orange) and in the presence of PH domain (green). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)
Superimposition of rigid docked form and eight IFD poses of 8 showed that IFD1, 2, 3, 4 and 7 had similar docked conformations and were different than rigid docked form, IFD8 and Inhibitor VIII (Fig. 6b). The Trp80 interacting bicyclic parts of rigid docked form and IFD8 were occupying the close region of Inhibitor VIII’s tricyclic interacting sub-structure, while IFD1, 2, 3, 4 and 7 were different than them, but showed good alignment between each other. IFD5 and IFD6 had completely different binding modes than other IFD models. The most significant side chain movement was observed for Trp80 and other moving side chains were of Lys268 and Arg273. The IFD results indicated that, favorable binding was observed for 1 and 8 with better GScore values than rigid docking and suggested a novel binding conformation for both ligands. 3.7. Selectivity analysis of proposed compounds 3.7.1. Docking of compounds selected by ligand and structure based pharmacophore models to kinase domain of Akt1 The compounds that were docked to allosteric binding site of Akt1 and had strain corrected GScores −9.0 kcal/mol and better (38 compounds from hits of RHH hypothesis and 158 compounds from hits of RAD hypothesis, 16 hits of APRRR hypothesis) were re-docked to a PH-domain excluded Akt structure with the aim of getting an idea on binding modes and affinities of the pre-docked ligands in the absence of PH domain of Akt1. An Akt1 structure that contains only kinase domain was selected as the receptor. Kinase domain of Akt1 (144–480, PDB code: 3QKL) was aligned to full length Akt1 (2–429, PDB code: 3O96) ˚ and inner (10 A) ˚ for selecting the same binding site. Outer (20 A) grid boxes were kept in same sizes and coordinates, with their center as the centroid of the native ligand (Inhibitor VIII, PDB code: IQO) of the Akt1 complex (PDB code: 3O96). A total of 212 compounds were docked to kinase domain without applying any constraints in Glide XP. The GScore values indicated that, the compounds that show very good binding affinity (with GScore values between −11.0 kcal/mol and −9.0 kcal/mol) when docked to Akt1 in the presence of PH domain, exhibited poor binding affinity (with GScore values between −5.5 kcal/mol and 6.5 kcal/mol) in the absence of PH domain. Compounds 1 from RAD-2 cluster, 8 from RAD-4 cluster and APRRR hit 20 were superimposed in the presence and absence of PH domain and were observed to change their binding modes completely (Fig. 7). These compounds were selected as representatives since they exhibited the highest straincorrected GScore values (1: −10.23 kcal/mol, 8: −10.17 kcal/mol, 20: −10.28 kcal/mol) in the presence of PH domain in addition to high fitness score to pharmacophore hypothesis (1: 1.98, 8: 2.22, 20: 1.8). The interaction diagrams of 12 compounds from
RAD-2 cluster, RAD-4 cluster and hits of APRRR hypothesis with the highest GScores were obtained to check the protein–ligand interactions. It was seen that the interactions with the Akt1 kinase domain residues in the presence of PH domain (i.e. Ser205, Lys268, Tyr 272, Asp 292) could not be observed for any of the ligands. Overall, different binding modes, higher binding energies and loss of key protein–ligand interactions suggest that, existence of PH domain is compulsory for allosteric inhibition by the proposed set of ligands. 3.7.2. PH domain sequence similarity BLAST algorithm (Altschul et al., 1990) was used to detect sequence similarity for PH domain of Akt1 (1–133) in Homo sapiens with the seven human kinases, PKD, PDK1, IRS1, SGK1, BTK, Akinase anchor protein and diacylglycerol kinase eta isoform 1, that have a PH domain. Sequence alignment results are shown in Fig. S4 and Table S10. PH domain of A-kinase anchor protein was detected to be most identical to Akt1 with a query cover (QC) percentage of 55% and maximum identity percentage (MI) of 50%. Correspondence of important residues of Akt1 on Akt2 and Akt3 isoforms and seven PH domain including kinases is shown in Table S11. The low similarity amongst PH domains and the low conservation in residues of interest suggest that targeting PH domain of Akt1 may result in higher selectivity between kinases. 4. Conclusion In this study, pharmacophore modeling and docking approaches were combined to identify a diverse set of scaffolds as potential allosteric inhibitors for Akt1. Both ligand-based and structurebased pharmacophore modeling were employed. The ligand-based pharmacophore model APRRR was based on previously identified allosteric inhibitors for Akt1 and structure-based pharmacophore models RHH and RAD used the information of allosteric binding site which was located between PH and kinase domains of Akt1. ZINC database was filtered separately based on 3D similarity to three pharmacophore hypotheses. Two-step docking (SP and XP) was performed using Glide protocol to predict binding modes and affinities of compounds in the pre-filtered database. Docking scores ranged between −12.2 and −0.5 kcal/mol and compounds with better GlideScore than −9.0 kcal/mol were subjected to post-docking evaluations. Hierarchical clustering, which was employed to determine two-dimensional similarities between top scoring hits, resulted in five clusters for a total of 196 hits of structure-based pharmacophore hypotheses RHH and RAD. Hits of ligand-based pharmacophore hypothesis APRRR were not clustered. Each cluster was further analyzed considering the existence of compulsory – stacking interaction with Trp 80 of Akt1, ADME and drug-likeness criteria and fitness to
12
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13
corresponding pharmacophore hypothesis. Induced-fit docking was performed for suggesting a favorable binding conformation for top scoring ligands. Common scaffolds for the surviving hits revealed three scaffolds as potential allosteric inhibitors for Akt1. Derivatives of; 3-methyl-xanthine, quinoline-4-carboxamide and 2-[4-(cyclohexa-1,3-dien-1-yl)-1H-pyrazol-3-yl]phenol were proposed for further studies as lead structures. Derivatives of methyl-xanthine were previously reported as human pancreatic lipase inhibitors, apnea reducers in preterm infants (HendersonSmart and Steer, 2001), therapeutics for severe exacerbations of chronic obstructive pulmonary disease (COPD) (Barr et al., 2003) and unique class of drugs for the treatment of asthma (Tilley, 2011). Derivatives of quinoline-carboxamide moieties were reported to enhance affinity of neurokinin 1 receptor antagonists, which are used as preventors for cancer chemotherapy side-effects such as vomiting (Cappelli et al., 2008). Additionally, quinoline4-carboxamide derivatives were reported to exhibit antimicrobial effects and influenced the growth of the microorganisms (Cappelli et al., 2008). This work provides a path for kinase and cancer researchers for further investigation of proposed scaffolds as potential drug candidates. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.compbiolchem. 2013.10.005. References Advanced Chemistry Development, Inc., Toronto, ON, Canada. Ajay, A., Walters, W.P., Murcko, M.A., 1998. Journal of Medicinal Chemistry 41, 3314–3324. Altschul, S.F., Gish, W., Miller, W., Myers, E.W., Lipman, D.J., 1990. Journal of Molecular Biology 215, 403–410. Arencibia, J.M., Pastor-Flores, D., Bauer, A.F., Schulze, J.O., Biondi, R.M., 2013. Biochimica et Biophysica Acta. Ashwell, M.A., Lapierre, J.M., Brassard, C., Bresciano, K., Bull, C., Cornell-Kennon, S., Eathiraj, S., France, D.S., Hall, T., Hill, J., Kelleher, E., Khanapurkar, S., Kizer, D., Koerner, S., Link, J., Liu, Y., Makhija, S., Moussa, M., Namdev, N., Nguyen, K., Nicewonger, R., Palma, R., Szwaya, J., Tandon, M., Uppalapati, U., Vensel, D., Volak, L.P., Volckova, E., Westlund, N., Wu, H., Yang, R.Y., Chan, T.C., 2012. Journal of Medicinal Chemistry 55, 5291–5310. Backman, T.W., Cao, Y., Girke, T., 2011. Nucleic Acids Research 39, W486–W491. Barnett, S.F., Defeo-Jones, D., Fu, S., Hancock, P.J., Haskell, K.M., Jones, R.E., Kahana, J.A., Kral, A.M., Leander, K., Lee, L.L., Malinowski, J., McAvoy, E.M., Nahas, D.D., Robinson, R.G., Huber, H.E., 2005. The Biochemical Journal 385, 399–408. Barr, R.G., Rowe, B.H., Camargo, C.A., 2003. Cochrane Database of Systematic Reviews (Online), CD002168. Bencsik, J.R., Xiao, D., Blake, J.F., Kallan, N.C., Mitchell, I.S., Spencer, K.L., Xu, R., Gloor, S.L., Martinson, M., Risom, T., Woessner, R.D., Dizon, F., Wu, W.I., Vigers, G.P., Brandhuber, B.J., Skelton, N.J., Prior, W.W., Murray, L.J., 2010. Bioorganic & Medicinal Chemistry Letters 20, 7037–7041. Blake, J.F., Xu, R., Bencsik, J.R., Xiao, D., Kallan, N.C., Schlachter, S., Mitchell, I.S., Spencer, K.L., Banka, A.L., Wallace, E.M., Gloor, S.L., Martinson, M., Woessner, R.D., Vigers, G.P., Brandhuber, B.J., Liang, J., Safina, B.S., Li, J., Zhang, B., Chabot, C., Do, S., Lee, L., Oeh, J., Sampath, D., Lee, B.B., Lin, K., Liederer, B.M., Skelton, N.J., 2012. Journal of Medicinal Chemistry 55, 8110–8127. Breitenlechner, C.B., Wegge, T., Berillon, L., Graul, K., Marzenell, K., Friebe, W.G., Thomas, U., Schumacher, R., Huber, R., Engh, R.A., Masjost, B., 2004. Journal of Medicinal Chemistry 47, 1375–1390. Calleja, V., Laguerre, M., Larijani, B., 2009a. Journal of Chemical Biology 2, 11–25. Calleja, V., Laguerre, M., Parker, P.J., Larijani, B., 2009b. PLoS Biology 7, e17. Cappelli, A., Giuliani, G., Anzini, M., Riitano, D., Giorgi, G., Vomero, S., 2008. Bioorganic & Medicinal Chemistry 16, 6850–6859. Cherrin, C., Haskell, K., Howell, B., Jones, R., Leander, K., Robinson, R., Watkins, A., Bilodeau, M., Hoffman, J., Sanderson, P., Hartman, G., Mahan, E., Prueksaritanont, T., Jiang, G., She, Q.B., Rosen, N., Sepp-Lorenzino, L., Defeo-Jones, D., Huber, H.E., 2010. Cancer Biology & Therapy 9, 493–503. ConfGen, Schrödinger, LLC, New York, NY, 2012. Dixon, S.L., Smondyrev, A.M., Knoll, E.H., Rao, S.N., Shaw, D.E., Friesner, R.A., 2006a. Journal of Computer-aided Molecular Design 20, 647–671. Dixon, S.L., Smondyrev, A.M., Rao, S.N., 2006b. Chemical Biology & Drug Design 67, 370–372. Du-Cuny, L., Song, Z., Moses, S., Powis, G., Mash, E.A., Meuillet, E.J., Zhang, S., 2009. Bioorganic & Medicinal Chemistry 17, 6983–6992.
Freeman-Cook, K.D., Autry, C., Borzillo, G., Gordon, D., Barbacci-Tobin, E., Bernardo, V., Briere, D., Clark, T., Corbett, M., Jakubczak, J., Kakar, S., Knauth, E., Lippa, B., Luzzio, M.J., Mansour, M., Martinelli, G., Marx, M., Nelson, K., Pandit, J., Rajamohan, F., Robinson, S., Subramanyam, C., Wei, L., Wythes, M., Morris, J., 2010. Journal of Medicinal Chemistry 53, 4615–4622. Friesner, R.A., Banks, J.L., Murphy, R.B., Halgren, T.A., Klicic, J.J., Mainz, D.T., Repasky, M.P., Knoll, E.H., Shelley, M., Perry, J.K., Shaw, D.E., Francis, P., Shenkin, P.S., 2004. Journal of Medicinal Chemistry 47, 1739–1749. Garofalo, R.S., Orena, S.J., Rafidi, K., Torchia, A.J., Stock, J.L., Hildebrandt, A.L., Coskran, T., Black, S.C., Brees, D.J., Wicks, J.R., McNeish, J.D., Coleman, K.G., 2003. The Journal of Clinical Investigation 112, 197–208. Garuti, L., Roberti, M., Bottegoni, G., 2010. Current Medicinal Chemistry 17, 2804–2821. Glide, Schrödinger, LLC, New York, NY, 2012. Green, C.J., Goransson, O., Kular, G.S., Leslie, N.R., Gray, A., Alessi, D.R., Sakamoto, K., Hundal, H.S., 2008. The Journal of Biological Chemistry 283, 27653–27667. Hahne, J.C., Meyer, S.R., Gambaryan, S., Walter, U., Dietl, J., Engel, J.B., Honig, A., 2013. International Journal of Oncology 42, 1630–1635. Hartnett, J.C., Barnett, S.F., Bilodeau, M.T., Defeo-Jones, D., Hartman, G.D., Huber, H.E., Jones, R.E., Kral, A.M., Robinson, R.G., Wu, Z., 2008. Bioorganic & Medicinal Chemistry Letters 18, 2194–2197. Henderson-Smart, D.J., Steer, P., 2001. Cochrane Database of Systematic Reviews (Online), CD000140. Hennessy, B.T., Smith, D.L., Ram, P.T., Lu, Y., Mills, G.B., 2005. Nature Reviews. Drug Discovery 4, 988–1004. Hers, I., Vincent, E.E., Tavare, J.M., 2011. Cellular Signalling 23, 1515–1527. Irwin, J.J., Sterling, T., Mysinger, M.M., Bolstad, E.S., Coleman, R.G., 2012. Journal of Chemical Information and Modeling. Kallan, N.C., Spencer, K.L., Blake, J.F., Xu, R., Heizer, J., Bencsik, J.R., Mitchell, I.S., Gloor, S.L., Martinson, M., Risom, T., Gross, S.D., Morales, T.H., Wu, W.I., Vigers, G.P., Brandhuber, B.J., Skelton, N.J., 2011. Bioorganic & Medicinal Chemistry Letters 21, 2410–2414. Kettle, J.G., Brown, S., Crafter, C., Davies, B.R., Dudley, P., Fairley, G., Faulder, P., Fillery, S., Greenwood, H., Hawkins, J., James, M., Johnson, K., Lane, C.D., Pass, M., Pink, J.H., Plant, H., Cosulich, S.C., 2012. Journal of Medicinal Chemistry 55, 1261–1273. LigPrep, Schrödinger, LLC, New York, NY, 2012. Lindsley, C.W., 2010. Current Topics in Medicinal Chemistry 10, 458–477. Lindsley, C.W., Zhao, Z., Leister, W.H., Robinson, R.G., Barnett, S.F., Defeo-Jones, D., Jones, R.E., Hartman, G.D., Huff, J.R., Huber, H.E., Duggan, M.E., 2005. Bioorganic & Medicinal Chemistry Letters 15, 761–764. Lipinski, C.A., Lombardo, F., Dominy, B.W., Feeney, P.J., 2001. Advanced Drug Delivery Reviews 46, 3–26. Liu, P., Cheng, H., Roberts, T.M., Zhao, J.J., 2009. Nature Reviews. Drug Discovery 8, 627–644. Loving, K., Salam, N.K., Sherman, W., 2009. Journal of Computer-aided Molecular Design 23, 541–554. Luo, Y., Shoemaker, A.R., Liu, X., Woods, K.W., Thomas, S.A., de Jong, R., Han, E.K., Li, T., Stoll, V.S., Powlas, J.A., Oleksijew, A., Mitten, M.J., Shi, Y., Guan, R., McGonigal, T.P., Klinghofer, V., Johnson, E.F., Leverson, J.D., Bouska, J.J., Mamo, M., Smith, R.A., Gramling-Evans, E.E., Zinker, B.A., Mika, A.K., Nguyen, P.T., Oltersdorf, T., Rosenberg, S.H., Li, Q., Giranda, V.L., 2005. Molecular Cancer Therapeutics 4, 977–986. Madhavi Sastry, G., Adzhigirey, M., Day, T., Annabhimoju, R., Sherman, W., 2013. Journal of Computer-aided Molecular Design 27, 221–234. Mahadevan, D., Powis, G., Mash, E.A., George, B., Gokhale, V.M., Zhang, S., Shakalya, K., Du-Cuny, L., Berggren, M., Ali, M.A., Jana, U., Ihle, N., Moses, S., Franklin, C., Narayan, S., Shirahatti, N., Meuillet, E.J., 2008. Molecular Cancer Therapeutics 7, 2621–2632. Milburn, C.C., Deak, M., Kelly, S.M., Price, N.C., Alessi, D.R., Van Aalten, D.M., 2003. The Biochemical Journal 375, 531–538. Moses, S.A., Ali, M.A., Zuohe, S., Du-Cuny, L., Zhou, L.L., Lemos, R., Ihle, N., Skillman, A.G., Zhang, S., Mash, E.A., Powis, G., Meuillet, E.J., 2009. Cancer Research 69, 5073–5081. Palmieri, L., Rastelli, G., 2013. Drug Discovery Today 18, 407–414. Phase, Schrödinger, LLC, New York, NY, 2012. Protein Preparation Wizard, Schrödinger, LLC, New York, NY, 2012. Prime, Schrödinger, LLC, New York, NY, 2012. QikProp, Schrödinger, LLC, New York, NY, 2012. Rhodes, N., Heerding, D.A., Duckett, D.R., Eberwein, D.J., Knick, V.B., Lansing, T.J., McConnell, R.T., Gilmer, T.M., Zhang, S.Y., Robell, K., Kahana, J.A., Geske, R.S., Kleymenova, E.V., Choudhry, A.E., Lai, Z., Leber, J.D., Minthorn, E.A., Strum, S.L., Wood, E.R., Huang, P.S., Copeland, R.A., Kumar, R., 2008. Cancer Research 68, 2366–2374. Saini, K.S., Loi, S., de Azambuja, E., Metzger-Filho, O., Saini, M.L., Ignatiadis, M., Dancey, J.E., Piccart-Gebhart, M.J., 2013. Cancer Treatment Reviews. Salam, N.K., Nuti, R., Sherman, W., 2009. Journal of Chemical Information and Modeling 49, 2356–2368. Schrödinger, LLC, New York, NY, 2012. Tilley, S., 2011. Methylxanthines in Asthma. In: Methylxanthines. Springer, Berlin, Heidelberg, pp. 439–456. Wang, P., Zhang, L., Hao, Q., Zhao, G., 2011. Mini Reviews in Medicinal Chemistry 11, 1093–1107. Watts, K.S., Dalal, P., Murphy, R.B., Sherman, W., Friesner, R.A., Shelley, J.C., 2010. Journal of Chemical Information and Modeling 50, 534–546.
O.G. Yilmaz et al. / Computational Biology and Chemistry 48 (2014) 1–13 Wu, Z., Robinson, R.G., Fu, S., Barnett, S.F., Defeo-Jones, D., Jones, R.E., Kral, A.M., Huber, H.E., Kohl, N.E., Hartman, G.D., Bilodeau, M.T., 2008a. Bioorganic & Medicinal Chemistry Letters 18, 2211–2214. Wu, Z., Hartnett, J.C., Neilson, L.A., Robinson, R.G., Fu, S., Barnett, S.F., Defeo-Jones, D., Jones, R.E., Kral, A.M., Huber, H.E., Hartman, G.D., Bilodeau, M.T., 2008b. Bioorganic & Medicinal Chemistry Letters 18, 1274–1279.
13
Wu, W.I., Voegtli, W.C., Sturgis, H.L., Dizon, F.P., Vigers, G.P., Brandhuber, B.J., 2010. PloS One 5, e12913. Yang, J., Cron, P., Good, V.M., Thompson, V., Hemmings, B.A., Barford, D., 2002. Nature Structural Biology 9, 940–944. Zhang, J.H., Lai, F.J., Chen, H., Luo, J., Zhang, R.Y., Bu, H.Q., Wang, Z.H., Lin, H.H., Lin, S.Z., 2013. Oncology Letters 5, 43–48.