Exploration for novel inhibitors showing back-to-front approach against VEGFR-2 kinase domain (4AG8) employing molecular docking mechanism and molecular dynamics simulations

A systematic literature survey was conducted to extract the dataset for VEGFR-2 inhibitors [30]. From the obtained information of 63 diverse compounds [31‐39], 24 structurally diverse compounds were referred to as training set and the remaining 39 were grouped into test set compounds. The training set compounds were employed to build the pharmacophore model while the test set was used to validate the same. The important criteria in choosing the training set compounds is the inclusion of the most active compounds into it such that they impart the most reliable information pertaining to the generated pharmacophore model. Additionally, the training set compounds exhibited a wide range of, half maximal inhibitory concentration, IC₅₀, spanning between 0.2 to 45,000 nmol/L. Further, the compounds in the training set, (Fig. 1), were designated as most active (IC₅₀ < 250 nmol/L, +++), moderately active (250 nmol/L ≤ IC₅₀ < 5000 nmol/L, ++) and inactive (IC₅₀ ≥ 5000 nmol/L, +), that was adapted based on the IC₅₀values. The same criteria were followed for the test set compounds. Their corresponding two dimensional (2D) structures were sketched using the ChemSketch (http://​www.​acdlabs.​com/​ resources/​ freeware/​ chemsketch/​ Advanced Chemistry Development (ACD) Inc., Toronto, Canada) and were translated to their three dimensional (3D) structures adapting the Discovery Studio v4.5 (DS).

In the recent times, pharmacophore modelling takes advantage of being one of the most reliable methods in identifying novel leads for different targets. For the present investigation, the three dimensional quantitative structure-activity relationship (3D–QSAR) based pharmacophore model was generated using the Catalyst HypoGen algorithm provided with the DS v4.5. This exploits the chemical features of the training set compounds and the conformation with the least energy were developed employing the BEST algorithm. In order to generate the best pharmacophore model, the energy and the uncertainty value were fixed at 20 kcal/mol and 3, respectively [40]. Further, Feature Mapping protocol was employed for investigating into the chemical features and to recognize the common features present in the training set that could be essential in the pharmacophore generation. This protocol has an ability to construct pharmacophore features available with the training set compounds and further these identified features play a critical role in the generation of the model. Amongst the generated models, the best hypothesis was chosen based upon the Debnath’s method [41].

With an aim to determine the predictive ability and its capability to identify the active compounds from that of the inactives, the selected pharmacophore was subjected to validation recruiting three different approaches such as, Fischer’s randomization, test set method, and the decoy set method. Fischer’s randomization was carried out alongside the pharmacophore generation, which prompts random spreadsheets based upon the selected level of confidence. For the present investigation, the confidence level was chosen to be 95%. The test and the decoy method of validations were conducted in order to understand if the generated pharmacophore was able to select the compounds in a similar manner as for the experimental activities. Ligand Pharmacophore Mapping protocol available on the DS was employed with Best Flexible algorithm. Test set was assembled with 39 structurally different compounds. The decoy set was instituted with a database of 710 compounds consisting of 15 active compounds. Following this, the enrichment factor (EF) and the goodness of fit score (GF) were computed using the formulae,

Here, Ha represents the total number of active compounds, Ht refers to total hits redeemed from the database, A refers to the total number of active compounds in the database, and D denotes the total number of molecules in the database.

Validated pharmacophore was employed as the 3D query to screen the databases such as Maybridge, Asinex, Chembridge, and NCI to retrieve novel scaffolds against angiogenesis, if the chemical compounds mapped with all the chemical features present in the pharmacophore. In this pursuit, the Ligand Pharmacophore Mapping protocol was used with Best–Flexible options.

Drug-likeness assessment was performed to the retrieved compounds from the databases so as to assess their biological activities. Accordingly, to judge the compound for strong pharmacokinetic properties, ADMET [42] and Lipinski’s rule were applied. ADMET specifically evaluates if the compound can cross the Blood Brain Barrier (BBB), allowable solubility, good intestinal absorption with less toxicity. Therefore, the values 3, 3 and 0 were fixed for BBB, solubility and absorption, correspondingly and were computed adapting ADMET Descriptors module on the DS. Additionally, the Lipinski’s Rule of 5 [43] was applied to the above filtered compounds to quantify if the prospective drug molecules could be well absorbed. This rule recommends a compound should have less than 10 hydrogen bond acceptors, less than 5 hydrogen bond donor groups having a molecular weight of less than 500 Da with log p value of less than 5 with 10 rotatable bonds. All the compounds that satisfied the aforementioned criteria were forwarded for the docking studies.

Challenging the potential screened lead molecules with the reliable drug target and to assess the degree of their binding affinities rendered in terms of the dock scores happens to be one of the most significant methodologies in drug discovery. Typically, this approach was deduced to assess the nature of the lead molecules in the active site and thereby its conformation. For the current study, Genetic Optimization for Ligand Docking v5.2.2 (GOLD) has been recruited [44, 45]. Target protein with the protein data bank (PDB) code, 4AG8, with high resolution of 1.95 Å, co-crystalled with axitinib was selected. Missing residues of the protein were rectified and all the hydrogen atoms were added, removing the water molecules [46]. Histidine tautomer orientations were placed in agreement with the crystal structure. The binding site of the protein was calculated for all the atoms that lie within the bound ligand around 15 Å. Furthermore, GOLD score was specified to understand the binding affinities between the ligand and the drug target, while the Chemscore was adapted for enumerating the rescoring function. Moreover, the GOLD score was initiated to generate 50 docking poses for each ligand and the reliable pose was selected based upon the highest dock score, molecular interactions and the hydrogen bonds that resulted between the ligand and the amino acid residues present at the active site of the protein molecule. Hereinafter, the most active compound in the training set was labeled as the reference compound.

The lead molecules identified after challenging against the 4AG8 would be further challenged with the CDK2 protein, 1URW, a potential target for cancer progression.

DFT is one of the most dynamic methods adapted to calculate the electronic structure of matter and thus provides the most valuable information with respect to the selected inhibitors. DFT for the resultant docking molecules was performed using Becke, 3-parameter, Lee-Yang Parr (B3LYP) [47], available on the DS in order to evaluate their orbital energies such as highest occupied molecular orbitals (HOMO) and lowest unoccupied molecular orbitals (LUMO). HOMO refers to the electron donor and LUMO denotes the electron acceptor. DFT was executed together with the Hit compounds and the compounds from the training set.

To gain further insight into the protein-ligand interactions, the procured Hit compounds from the docking studies and the DFT were subjected to the Molecular Dynamics (MD) simulations along with the reference compound. The ligand topologies were generated utilizing the SwissParam [48‐50], while the topologies of the protein were generated employing Chemistry at HARvard macromolecular Mechanics force-field (CHARMm ff) [51‐54] implemented in Groningen Machine for Chemical Simulations (GROMACS) 5.0.6 [55]. Dodecahedron box was obtained and was solvated with three-site transferable intermolecular potential (TIP3P) water model followed by neutralizing the system with the counter ions. All the bad contacts were further removed by subjecting the system to pass through steepest descent algorithm at 10,000 steps with an upper limit of the force being lower than 1000 kJ/mol [56]. Following this, the equilibration was conducted by Number of particles, Volume and Temperature (NVT) [57] and Number of particles, Pressure and Temperature (NPT) [58] at 100 ps at 300 k and 100 ps at a pressure of 1 bar maintained by Parrinello-Rahman barostat and allowing the movement of the counter ions and the water molecules, constraining the protein backbone. Linear Constraint Solver for Molecular Simulations (LINCS) [59] algorithm was used to restrain heavy atom bonds and their respective hydrogen atoms. Particle Mesh Ewald (PME) [60] was utilized to compute the long rage electrostatic interaction and a cut-off distance of 12 Å was attributed for Coulombic and van der Waals interactions. MD simulations were performed for 30 ns storing the coordinate data for every 2 fs. Corresponding results were evaluated employing the Visula Molecular Dynamics (VMD) [61] and DS, respectively.

Using the HypoGen algorithm provided with the DS, 24 training set compounds were employed to develop the pharmacophore model, (Figs. 1 and 2), which resulted in the generation of 10 hypotheses, Table 1, upon the utilization of 3D QSAR Pharmacophore Generation protocol available with the DS. The preferred features for the pharmacophore generation were hydrogen bond acceptor (HBA), hydrogen bond donor (HBD), hydrophobic (HyP), hydrophobic aliphatic (Hy-Ali) and ring aromatic (RA).

From the generated models, Hypo1 was chosen as the best pharmacophore model as it satisfied the Debnath’s rules, which states that a good pharmacophore model should consists of high cost difference, good correlation, least RMSD and low cost values. The generated pharmacophore model composed of aromatic feature (RA), one hydrophobic aliphatic feature (Hy-Ali) and two hydrophobic (HyP) features.

Further, to evaluate the predictive ability of the Hypo1, it was allowed to examine the inhibitory activities of 24 training set compounds. The training set compounds were grouped into most active, moderately active and the least active compounds based upon their IC₅₀ values as, (IC₅₀ < 250 nmol/L, +++), (250 nmol/L ≤ IC₅₀ < 5000 nmol/L, ++) (IC₅₀ ≥ 5000 nmol/L, +), correspondingly. Hypo1 calculated the inhibitory activity values of the training set in accordance with the experimental values, Table 2. Furthermore, Hypo1 has successfully mapped to the most active compound and the most inactive compound, (Figs. 3 and 4).

In order to assess the quality of the generated pharmacophore, it was subjected to a series of validations such as, Fischer’s randomization method, test set method and the decoy test method.

Validated Hypo1 was employed to screen and retrieve the novel compounds against the VEGFR-2. The chemical features imbibed within the Hypo1 display an important role in identifying the new leads. Four databases, such as Chembridge, Maybridge, Asinex and NCI were selected for screening the dual inhibitors. Accordingly, Hypo1 has successfully mapped with 12,080, 14,521, 29,836 and 29,660 compounds from the aforementioned databases. Following which, compounds with a greater fit value than 8 were proceeded to the Lipinski’s Rule of Five and the ADMET studies. These two studies quantify the pharmacokinetics of a drug molecule and thus, is an essentiality for a drug molecule to qualify them. ADMET particularly computes BBB penetration, solubility, human intestinal absorption (HIA) solubility, plasma protein binding (PPB) and CytochromesP450 (CYPs) inhibition. The results were evaluated by setting the standard as 3, 3, and 0 for solubility, BBB and absorption, correspondingly. From the four databases, a total of 699 molecules were identified possessing the drug-like properties. The screened compounds along with the training set compounds were subjected to molecular docking analysis. The overall process of screening is represented pictorial form (Fig. 7).

Genetic Optimization for Ligand Docking, (GOLD) v5.2.2 was recruited for performing the molecular docking assay. GOLD has an ability to conduct lead optimization and to determine the accurate binding pose analysis. Technically, GOLD operates by two scoring functions, Goldscore fitness, and the Chemscore. Goldscore is a fitness parameter that is governed by four components, such as; protein-ligand hydrogen bond energy (external H-bond), protein-ligand van der Waals (vdw) energy (external vdw), ligand internal vdw energy (internal vdw), ligand torsional strain energy (internal torsion), respectively. This score was optimized for understanding and predicting the position of the bound ligand. On the other hand, Chemscore was used for rescoring and further computes the total free energy change on the ligand binding. The aptness of the docking parameters were assessed by subjecting the co-crystal to dock into the selected active site. This generated a reasonable RMSD of 0.9 Å between the docked pose and the co-crystal thus ensuring the accuracy of the GOLD parameters. Therefore, these parameters were considered for docking the screened compounds into the active site of the target, (Additional file 2).

Docking of the screened ligands and the training set compounds was initiated after specifying the radius of the active site around the co-crystal to 15 Angstrom (Å) followed by allowing the generation of 50 conformations for each ligand. All those compounds which fall in the selected radius were considered for the succeeding steps. The docking studies were initially executed with 4AG8, a potential drug target for angiogenesis. The final selected lead compounds after the DFT and the MD simulations were challenged with the CDK2 protein, 1URW. This strategy was conceived to identify a common drug molecule for both angiogenesis and progression.

The angiogenesis target for the present study was 4AG8, retrieved for the Protein Data Bank (PDB, http://​www.​rcsb.​org/​pdb/​home/​home.​do). The reference molecule has displayed a dock score of 89.77 and therefore, only those lead molecules whose scores were above the reference molecules were chosen for further investigation and thus, qualifying 18 screened molecules. These compounds were examined manually for the hydrogen bond interactions with the amino acids residues, Cys919, Glu917, Asp1046, and Glu885, respectively. Out of 18 screened lead molecules, only 11 exhibited the hydrogen bond interactions with the key residues. These 11 compounds, (Fig. 8), obeyed all the filters employed to identify an efficient lead candidate. These 11 molecules were subjected to the Density Functional Theory (DFT) analysis for further understanding their molecular orbital energies.

Highest occupied molecular orbital (HOMO) and the lowest unoccupied molecular orbital (LUMO) that ascribe the DFT were computed with six training set compounds (2 active, 2 moderately active and 2 inactive compounds) and 11 screened compounds. Small gap obtained, reports a compound to be highly reactive, while the larger gap signifies the presence of low reactivity of the compound to that of the target protein [62]. Accordingly, compound 4 has been selected as it demonstrated the least gap, Table 4, obeys the back to front approach and maps to all the pharmacophore features, (Fig. 9). Hereafter, is compound was named Hit.

With an objective of assessing the binding stability of the final systems, the best conformations obtained from the docking were proceeded to MD simulations. Furthermore, MD simulation evaluates and delineates on their dynamic behaviour and with each other. The MD results were examined for the RMSD values, potential energy and the radius of gyration to assess the stability of the protein backbone. The RMSD values were found to be 1.0 nm, and 0.8 nm, respectively for the reference molecule and the Hit. Additionally, it can be observed that the RMSD profiles of the candidate compound was more stable than the reference, (Fig. 10). It was also noticed that the stability was attained after 20 ps for reference compound and while Hit 1 was stable after 25 ps, (Fig. 10). Furthermore, their stability was assessed by plotting the potential energy, (Fig. 11) and the radius of gyration, (Fig. 12). Both profiles have rendered results that showed no abnormal behaviour throughout the simulation. Furthermore, the protein-Hit complex was found to be more compact as compared with the protein-reference complex, (Fig. 12). Additionally, from the Fig. 12 it can be understood that there were a few aberrations exhibited by the protein- reference complex and was stable after 25 ns. On the contrary, the protein-Hit complex was found to be stable after 16 ns, showing no aberrations thereafter.

The binding mode analysis was performed retrieving the representative structures from the last 3 ns trajectories. Upon superimposition, it was observed that the binding pattern of the Hit was similar with the reference compound, (Fig. 13). Further inspecting the molecular interactions of the reference, it was revealed that the reference has formed three hydrogen bonds demonstrated by two key residues, Cys919 and Asp1046, respectively and displaying an acceptable bond length, Table 5. The HN atom of the Cys919 interacted with N7 atom of the ligand and NH and O atoms of Asp1046 have joined with O26 and H62 of the ligand, (Fig. 14a). Leu840 has participated in the π-sigma interaction, while the residues, Ala866, Ala866, Val898, Leu889, and Leu1035 were found to interact through the π-alkyl bond, (Additional file 3). The Hit has interacted with the protein target with three hydrogen bonds represented one each by Glu885, Cys919 and Asp1046, respectively, (Fig. 14b). HN atom of Cys919 has involved with the ligand’s O14 atom with a hydrogen bond length of 2.16 Å and the OE2 atom of Glu885 joined to the OE2 atom of the ligand displaying a bond length of 2.09 Å, respectively. Additionally, Leu840, Val848, Lys868, Leu1035 have interacted involving with the π-π bonds. Leu889 and Phe918 have interacted with the π-alkyl bond and thus make the ligand to be seated firmly within the active site groove, (Additional file 4). Moreover, from the large set of screened databases only one Hit was identified has the only retrieved compound that obeyed back-to-front pattern of binding and thus projecting itself as a unique lead compound, (Fig. 15). Correspondingly, it was observed that the Hit molecule has been positioned within the DFG motif comprising of Asp1046-Phe1047-Gly1048 residues sandwiched between Asp1046 and Glu885 on either side and hence crucial in conferring the reduction of angiogenesis. The Cys1045 has formed a π- sulfur bond that makes the ligand to be placed firmly with the active site. A hydrophobic interaction have additionally concentrated towards the DFG motif of the activation loop and thereby, distorts the allosteric change of the motif and inhibits the kinase activity [5] and therefore the Hit may act as type II inhibitor. Further details of the interactions are tabulated in Table 5. Furthermore, to gain insight into and attain a proper reasoning on the nature of the hydrogen bonds in the active site, the hydrogen bonds were recorded throughout the simulation time. Consequently, it was noticed that the reference has produced an average of 0.7 hydrogen bonds, while the Hit has displayed a higher average of 1.7, (Fig. 16).

After achieving the first objective, now the investigation proceeds to understand if the ligand is potential enough to act against cancer prognosis. Cyclin-Dependent Kinase 2(CDK2) has been attributed with certain cancer progressions, especially the oral cancer [63, 64]. Progression refers to the advancement of the disease during the course of the time. For the current study the protein 1URW, was retrieved from the protein data bank. This protein is in complex with imidazole [1,2-B] pyridazine inhibitor of 1.6 Å. The identified lead molecule form the above steps is now challenged with 1URW, such that a common drug could be identified.

All the steps, such as, protein preparation and ligand preparation were followed as per the aforementioned methods. The active site of the protein was plotted at 12 Å around the inbuilt co-crystal. To ensure the suitability of the docking parameters, the co-crystal was initially docked and an acceptable RMSD of 1.75 Å was generated, (Additional file 5) and the same parameters were considered for docking employing GOLD v5.2.2. The results were examined taking into consideration that the inhibitor lies within the selected active site sphere and forms a hydrogen bond with Leu83, Asp86 and Lys89. Amongst them, Leu83 is the core residue as reported earlier [26, 65]. To authenticate the binding of the ligand with the active site residues, the results generated by the co-crystal docking were taken into consideration. The best pose generated was escalated to MD simulation studies to assess the protein backbone stability and were read as root mean square deviation (RMSD). Subsequent results demonstrated that the cocrystal, reference and the Hit have displayed stable RMSD values that exist below 0.25 nm throughout 30 ns run, (Fig. 17).

The representative structures from the last 3 ns were extracted and superimposed for further analysis. Upon superimposition of the co-crystal, reference and the Hit have demonstrated a similar type of binding mode, (Fig. 18). The dock results revealed that the co-crystal has formed four hydrogen bonds with residues namely Leu83, Asp86 and Lys89 demonstrating a dock score of 53.05, while the reference and Hit has displayed a score of 52.56 and 64.34. Additionally, a number of hydrophobic bonds have been generated in the form of van der Waals interactions and π-π bonds as represented in Fig. 19a and Additional file 6. The reference compound on the other hand has rendered five hydrogen bonds one each by Glu12 and Gly16 and three bonds with Lys89. It was observed that only the Lys89 residue was noticed to participate as was seen in the crystal structure. Additionally, Glu12 formed a weak hydrogen bond of 3.0 Å and displayed no interaction with Leu83. This could be because the compound is a specific VEGFR-2 inhibitor. However, it exerts its effect through the involvement of Lys89 residue, (Fig. 19b: Additional file 7). The retrieved Hit has displayed three hydrogen bonds involving two crucial residues, Leu83 and Lys89 and further has generated an acceptable bond length of 2.1 Å and 2.4 Å, respectively, Fig. 19C. Additionally, it formed more number of hydrophobic interactions with the key resides, (Additional file 8). The details of the molecular interactions are tabulated in Table 6.

Furthermore, the binding pockets of both the targets were evaluated to understand their similarity and was assessed using PocketMatch (PM) [66] that was executed through three aspects (a) representation of each site as sorted lists of distances between chosen points, (b) alignment of two sets of distance lists and (c) choosing a scoring scheme for arriving at a final score. The similarity is read as PM score which ranges from 0 to 1, where zero implies no similarity and 1 refers to complete similarity. The obtained results revealed the PM score to be 0.82 (0.6 or greater refers to almost similar). To further affirm the similarity of the active sites, the align structures protocol available on the DS was initiated and subsequently, the arrangement of the sequences, (Fig. 20) and the active site residues, (Fig. 21: Additional file 9) were evaluated. It was interesting to note that the majority of the key residues were conserved besides; the key residues Cys919 and Leu83 were complementary with each other and were instrumental in rendering the inhibitory activity in the corresponding drug targets. These findings further reinforce that they share similar binding pocket facilitating the use of common drug.