Introduction

The process by which new blood vessel is formed—dubbed Angiogenesis—has recently gained significant attention in anticancer drug development. The process of angiogenesis is pivotal in solid tumor growth, diabetes and other age-related diseases [1, 2]. In solid tumor, blood vessels are formed excessively to confer sufficient oxygen and nutrient supply to malignant cells in order to establish the process of metastasis [3]. Evidence has also shown that vascular endothelial growth factor (VEGF) and its receptors (VEGFRs) are the main modulators of angiogenesis [4, 5].

The members of VEGFR including VEGFRs 1–3 are targeted by VEGF, which binds to the extracellular domain of VEGFRs to activate downstream signaling pathways [3]. VEGFR-2 (KDR) directly regulates endothelial cell proliferation and migration [6,7,8]. Upon binding of VEGF-A, the tyrosine residues at the carboxy-terminal of the receptor are autophosphorylated, thus activating cell signaling, vasculogenesis, and angiogenesis [9, 10]. Aggressive expression of VEGFR-2 activates downstream pathways of cell survival, which makes this an attractive target for developing novel anticancer agents.

A number of VEGFR inhibitors are currently in phase I or phase II clinical trials [11], while other inhibitors, such as sorafenib (Nexavar®, BAY 43–9006), sunitinib malate (Sutent®, SU11248), and bevacizumab (Avastin®), are already approved for cancer treatment by the Food and Drug Administration (FDA) [12]. Pharmacologically, they act by blocking the VEGFR pathway; however, their prolonged use can lead to serious side-effects, such as bleeding, gastrointestinal perforation and hypertension [13, 14]. Therefore, the search for novel and potent VEGFR inhibitors that could effectively block angiogenesis remains a major challenge in anti-angiogenic drug development.

Due to their high efficacy and lower toxicity, compounds from natural sources have recently gained more attention as anti-angiogenesis agents [15]. Nevertheless, they still suffer from poor bioavailability [16]. Consequently, in vivo metabolites of some herbal compounds were found to show better pharmacokinetic, potency and safety profiles in comparison to their parent compounds [17]. For this reason, to find potent VEGFR-2 inhibitors with optimum pharmacokinetic profiles, this study applied a virtual in silico screening approach using a database of in vivo metabolites of herbal ingredients. The virtual screening procedure was integrated with a three-step filtering approach, including Glide XP docking, ADME/T and MM-GBSA analysis, to sort out the top hits from the database (Fig. S1, see Supplementary file 1). After that, each docked protein–ligand complex was subjected to 100-ns MD simulations along with binding energy calculations through MM-PBSA analysis, in order to elucidate the VEGFR-2 binding mechanisms of identified compounds. The study identified two virtual hits, ZINC14762520 and ZINC36470466, that could be potent VEGFR-2 inhibitors, and also suggested a basis for further rational design of their various analogs to find more effective inhibitors in the future.

Materials and methods

Database processing

A dataset of in vivo secondary metabolites of herbal ingredients was retrieved from the ZINC server named Herbal Ingredients In-Vivo Metabolism (HIM) (last updated: 12 August 2013), giving a set of 663 in-vivo metabolites of herbal active ingredients. Although the ZINC database consists of molecules ready to dock in three dimensional (3D) formats, to make it more rationale, we further prepared the dataset using Schrödinger Suite 2013 [18, 19] software. Here, all ligands were minimized in Ligprep2.5 wizard [20], by implementing OPLS 2005 [21] force field. During the minimization, the module Epik 2.2 [22] was utilized to fix the ionization state of the ligand at pH 7.0 ± 2.0 [23]. From this analysis, up to 32 possible stereoisomers of each compound were generated; we selected only the conformer having the lowest energy for further processing.

Protein preparation and grid generation

Protein structures for virtual screening were prepared using the protein preparation module of Schrödinger Suite 2013 [19] software, in which the crystal structure is initially assigned proper hydrogen, charges and bond orders. As the main target of the virtual screening was VEGFR-2, its tyrosine kinase domain was retrieved from the RCSB protein data bank (PDB id: 3B8R) [24]. At neutral pH, all hydrogen bonds in the structure were optimized, deleting unnecessary water. Afterwards, the minimization process was run with the OPLS 2005 force field, considering structural changes of not more than 0.30 Å of RMSD. The active site of the protein was fixed for docking simulation by generating a grid box at the reference ligand binding the protein. Grid generation parameters were kept at default, with a box size of 18 Å × 18 Å × 18 Å, and the OPLS 2005 force field utilized for post minimization. The charge cutoff and van der Waals scaling factor were set to 0.25 and 1.00, respectively [25].

Virtual screening

We carried out extra precision (XP) flexible docking using the Glide module of Schrödinger-Maestro v9.4 [26, 27], which is more sophisticated than SP/HTVS in scoring function [27]. Here, all ligands were treated flexibly, considering the partial charge and van der Waals factor of 0.15 and 0.80, respectively. Minimization was performed to the docked complex after docking using the OPLS 2005 force field. The best-docked pose with lowest Glide score value was recorded for each ligand [25].

ADME/T analysis

In order to predict the pharmacokinetic profile of the identified hits, QikProp (Schrodinger Suite 2013) [28] was utilized, where the absorption, distribution and metabolism of each ligand were calculated. Pharmacokinetics and the drug-like properties of the ligands, such as the rate of oral absorption, QlogP, solvent accessible surface area (SASA), and Lipinski’s rule of five [29], were predicted, and maximum percentage of oral absorption rate with a cutoff of 50% was considered for selecting the top hits for further processing.

Prime MM-GBSA

Binding affinity calculations were carried out using the Prime MM-GBSA module of the Schrodinger suite [30], where greater negative affinity denotes higher stability. The docked Pose viewer file from Glide XP docking was introduced for calculation, using the sampling minimization protocol used Generalized Born Surface Accessible (GBSA) as a continuum model, and OPLS 2005 force field as molecular mechanics (MM), keeping the protein flexible [31,32,33,34,35]. For correcting empirical functions of π-stacking and H-bond interactions, a dielectric solvent model such as VSGB 2.0 [36] was used.

MD simulations

MD simulations were performed using YASARA Dynamics software [37], and the AMBER14 force field [38]. A total of six different systems were used to run MD simulation. The systems included top four virtual hits, (ZINC13370550-VEGFR-2, ZINC02585769-VEGFR-2, ZINC14762520-VEGFR-2 and ZINC36470466-VEGFR-2 complexes), sunitinib-VEGFR-2, and apo-VEGFR-2. In each protein–ligand complex, the parameters of each ligand were assigned through AutoSMILES [39] algorithms, where unknown organic molecules are parameterized fully automatically by the calculation of semi-empirical AM1 Mulliken point charges [39] with the COSMO solvation model, assigning of AM1BCC [40] atom and bond types, and also assigning GAFF [41] (General AMBER Force Field) atom types and the remaining force field parameters. Before simulation, the hydrogen bonding network of the protein–ligand complex was optimized and solvated by a TIP3P [42] water model in a simulation cell. Periodic boundary conditions were maintained with a solvent density of 0.997 gL−1. Titratable amino acids in the protein complex were subjected to pKa calculation during solvation. The initial energy minimization process of each simulation system, consisting of 62,521 ± 10 atoms was performed by a simulated annealing method, using the steepest gradient approach (5000 cycles). Each simulation was run with a multiple time step algorithm [43], using a time-step interval of 2.50 fs under physiological conditions (298 K, pH 7.4, 0.9% NaCl) [44]. All bond lengths were constrained using the linear constraint solver (LINCS) [45] algorithm, and SETTLE [46] was used for water molecules. Long-range electrostatic interactions were described by the PME [47] methods, and, finally, 100 ns MD simulation was accomplished at Berendsen thermostat [48] and constant pressure. For further analysis, the trajectories were saved every 250 ps, and subsequent analysis was performed by default script of YASARA [49] macro and VMD [50, 51] software.

All snapshots were further subjected to MM-PBSA (MM-Poisson–Boltzmann surface area) binding free energy calculation by YASARA software using the following formula [52],

$$ \mathrm{Binding}\ \mathrm{Energy}={\mathrm{E}}_{\mathrm{potRecept}}+{\mathrm{E}}_{\mathrm{solvRecept}}+{\mathrm{E}}_{\mathrm{potLigand}}+{\mathrm{E}}_{\mathrm{solvLigand}}-{\mathrm{E}}_{\mathrm{potComplex}}-{\mathrm{E}}_{\mathrm{solvComplex}} $$

Here, YASARA built-in macros was used to calculate MM-PBSA binding energy, using AMBER 14 as a force field, where more positive energies indicate better binding [53].

Results and discussion

Virtual screening

Overexpression of VEGF and its receptor VEGFR-2 is the most common phenomenon in angiogenesis and plays a significant role in cell survival and promoting tumorgenesis [54,55,56]. In current drug discovery processes, virtual screening is the most widely used approach to find novel hits for a target by reducing time, cost and the massive compound library. Apparently, XP docking assists in filtering out false positives from the compound database by using a strict scoring function; hence, it was considered in this study [57]. However, prior to docking simulations, we first validated docking accuracy through redocking of a known reference ligand using XP docking methodology. Here, the reference co-crystal ligand was first separated and re-docked using the Glide XP algorithm. Conformation of the complex with the highest negative docking score was selected from each run. The RMSD of the docked conformation was then calculated, where a lower RMSD value denotes a high level of the accuracy of docking protocol. Interestingly, Glide docking with XP setting produced the docked complex with lowest RMSD of 0.61 Å, which denoted that the position of docked ligand was almost similar to that of the crystal, as shown in Fig. S2 (see supplementary file 1). Therefore, the candidate ligands including reference inhibitors, sunitinib and sorafenib were docked into VEGFR-2, following the same docking protocol.

Initially, the top 5% compounds were selected from Glide XP docking based on high docking score. As a result, 35 compounds were considered, with docking score range from −12.00 to −10.00 kcal mol−1, in which the highest docking score was found to be −11.80 kcal mol−1 for ZINC95099443 and the lowest docking score was −10.04 kcal mol−1 for ZINC95099251. On the contrary, the reference inhibitors, sunitinib and sorafenib resulted in docking scores of −6.60 kcal mol−1, and − 7.65  kcal mol−1, respectively (Table 1). However, bioavailability and toxicity are the primary concerns for drug development; therefore, we subjected these compounds for ADME/T profiling, as discussed below.

Table 1 Top six candidates generated from structure based virtual screening on the basis of extra precision (XP) Glide docking against VEGFR-2
Full size table

Bioavailability of ligands

Pharmaceutically relevant attributes and physiochemical descriptors of the identified hypothetical hits were calculated by QikProp software. Each descriptor’s range satisfies 95% of known drugs [58,59,60]. The pharmacokinetic properties of 35 ligands were assessed (see supplementary file 2), where cell permeability, the number of rotatable bonds, number of metabolites, solubility, logP, and oral absorption rate were predicted, from which the top six candidates were selected (Fig. S3, see supplementary file 1).

As described in Table 2, all calculated attributes were in the satisfactory range, with properties including, QPlogPo/w (octanol/water partition coefficient), QPlogHERG (log IC50 value for blockage of K+ channels) [61, 62], SASA [59], QPlogS (aqueous solubility) [59], molecular weight, and H bond acceptor/donor [61]. Although the standard inhibitors obtained highest oral absorption, they were also predicted to fall outside the recommended range for QPlogHERG value. The HERG K+ channel, which is encoded by the human ether-a-go-go related gene (HERG), is one of the main reasons for the cardiac toxicity produced by a wide range of therapeutic drugs [63]. This channel promotes electrical activity of the heart and also modulates various functions of the nervous system [64]; thus, blockers of this channel cause potential toxicity, which is a major consideration in drug discovery processes [65]. All these findings indicating that the ligands identified showing ideal behavior of drug molecules, and that they are well tolerated and less toxic than standard inhibitors.

Table 2 Pharmacokinetic and toxicological properties of the top six candidates and standard inhibitors, calculated from QikProp
Full size table

MM-GBSA binding affinity estimation

The top six identified ligands were re-ranked by their binding affinity calculated through MM-GBSA analysis: ZINC14762520 obtained highest binding affinity (ΔGBind = −78.09 kcal mol−1) with highest contribution from covalent and liphophilic binding energies. In contrast, ZINC95099475 showed the lowest binding affinity of −58.29 kcal mol−1 (Table S1, see supplementary file 1). In comparison to the identified ligands, the reference inhibitors sunitinib and sorafenib showed binding affinities of −45.26 kcal mol−1 and − 57.44 kcal mol−1 respectively. On the basis of MM-GBSA result, we thus selected the top four compounds and subjected them to detailed binding analysis.

Ligand binding analysis

The docked complexes of the top four virtual hits from Glide XP docking were further considered for protein–ligand interaction profiling. As shown in Table 3, and Fig. 1a, ZINC02585769 showed four hydrogen bonds with C919, D1046, E885 and G922 residues, while C919 formed double hydrogen bonds with the ligand, maintaining corresponding bonding distances of 2.23 Å and 2.41 Å, respectively. Although the ligand formed major hydrophobic interactions with K868, V899, V848, L840, A866 and L1035 through pi-alkyl bonding, C919 and K920 showed additional amide-pi stacking interactions with the ligand.

Table 3 Detailed molecular interactions analysis for best hits obtained from Glide XP docking
Full size table
Fig. 1a–d
figure 1

Binding modes of compoundswith VEGFR-2. Molecular interactions are displayed in hydrophobic surface view. Green dotted line Hydrogen bonds, pink dotted lines π-alkyl interactions. The hydrophobic intensities of the binding site ranged from −3.00 (least hydrophobic area—shaded blue) to 3.00 (highly hydrophobic area—shaded brown). a ZINC02585769, b ZINC13370550, c ZINC14762520, d ZINC36470466

Full size image

On the other hand, ZINC13370550 (Fig. 1b) formed several hydrogen bonds with the active site residues, where most of the interactions were mediated by E868, C919 and N923 residue. C919 was also involved in π-donor hydrogen bonding with the ligand, while L840, L1035, V848, A866 and C1045 were responsible for hydrophobic interactions via π-alkyl bonding.

In the case of the ZINC14762520-VEGFR-2 complex, ZINC14762520 showed hydrogen bonding to the active site, where the interactions were maintained by C919, D1046 and F1047 residues. Furthermore, ZINC14762520 also formed pi-alkyl bonds with K868, V899, A866, C919 and L1035 residues and were thus involved in hydrophobic interactions (Fig. 1c).

ZINC36470466 (Fig. 1d) formed a stable complex with VEGFR-2 by forming hydrogen bonds with T916, D1046 and C919 side chains, where both D1046 and C919 showed two hydrogen bonds individually. Here, V899, L840, V848, A866 and L1035 were involved in hydrophobic interactions by π-alkyl interactions.

According to the detailed analysis of molecular interactions, except ZINC14762520, all compounds formed hydrophobic interactions with L840 and L848 residues of the glycine rich loop (G-loop) that makes a hydrophobic interaction with ATP [66, 67], while ZINC14762520 showed interactions with several residues from the hinge region and DFG motif, including C919, D1046 and F1047. All ligands also made contacts with the residues of the linker region, including C1045 and D1046, which confirms their VEGFR-2 inhibition, as these residues are directly involved in polar interactions with ATP, assist the shifting of γ-phosphate through coordinating Mg2+ ions [68]. Furthermore, non-covalent interactions of the residues of the hinge region (T917 to N923) with all four ligands revealed that these compounds may act as ATP competitive inhibitors [69, 70].

MD simulations

To describe the binding mechanisms, flexibilities and structural behaviors of the top virtual hits, we conducted MD simulations for 100 ns for each complex, including the reference inhibitor, sunitinib and also the ligand free protein. Different energies such as coloumb, van der Waals and total energies were calculated for each system (summarized in Table S2). The atomic RMSDs of the Cα, C, and N atoms of protein in all complexes were calculated from the simulation trajectories and are plotted in Fig. 2. As shown in Fig. 2, the highest RMSD values were observed for ZINC36470466, while ZINC02585769 obtained the lowest RMSDs in the 100 ns simulation. As can be seen from the plot, ZINC02585769 showed a more rigid conformation than the other proteins, and also achieved equilibrium at 10 ns and remained stable afterward.

Fig. 2
figure 2

Time evolution of root-mean-square deviations (RMSDs) of backbone atoms (C, Cα, and N) for protein of each docked complex. Complexes: Blue apo, orange sunitinib, red ZINC02585769, olive green ZINC13370550, black ZINC14762520, pink ZINC36470466

Full size image

In contrast, ZINC36470466 showed a dramatic increase in flexibility, with RMSD values rising gradually from 1.5 Å to 2.75 Å over time. ZINC14762520 was in equilibrium up to 70 ns, but showed fluctuations thereafter. On the other hand, the apo form also reached at equilibrium after 10 ns and maintained RMSD up to 1.9 Å for 50 ns, after which the RMSD was seen to increase gradually. Similarly, sunitinib displayed stable RMSD up to 50 ns; however, after that it decreased to 1.5 Å and remained steady thereafter. Overall, the results of RMSD analysis clearly depicted that the conformation of the protein was changed upon ligand binding, as the apo structure represented the lowest RMSD fluctuations over the course of the simulation. In order to investigate local fluctuations in protein structure induced by the ligands, the root mean square fluctuation (RMSF) of each protein backbone was calculated. As shown in Fig. 3a, five flexible regions were involved in the ATP binding site, including G-loop L836 to I849, hyd-1 A866 to K868, hinge T916 to N923, catalytic subunit H1026 to L1035 and DFG motif C1045 to G1048. Among these compounds, ZINC36470466 produced the highest fluctuations in G-loop regions, while ZINC14762520 showed low RMSF compared to the apo form. Again in all cases, amino acids ranging from 114 to 128 fluctuated more in apo form; however, fluctuations were seen to reduce after ligand binding. In the case of the standard inhibitor, sunitinib increased flexibility in the hinge region, up to 1.9 Å, while no significant changes in RMSF were observed for ZINC14762520. It is also noteworthy that binding of all ligands except ZINC14762520 increased local fluctuations of the catalytic subunit region in comparison to the apo protein. This was due to their strong interactions with residues and rotation of the particular pharmacophore involved in intermolecular interactions. In the DFG motif, ZINC13370550 produced the highest RMSF, which denotes increased flexibility in this region. Overall, RMSF analysis suggested that all compounds formed favorable interactions with ligand binding regions. Although dynamic changes occurred upon binding to these ligands, protein stability was increased, which is similar to the reference ligand.

Fig. 3
figure 3

a Structural representation of VEGFR-2 kinase highlighting the different motifs involved in ATP binding. b Root mean square fluctuation (RMSF) plots of protein–ligand complexes: Blue apo, orange sunitinib, red ZINC02585769, olive green ZINC13370550, black ZINC14762520, pink ZINC36470466. c Comparative RMSF plot between protein-ligand complexes and apo structure. Residues involved: light green G-loop, orange hyd-1, blue hinge, brick red catalytic subunit, cyan DFG motif

Full size image

To further clarify the structural deviation derived by the ligands, flexibilities of the protein in terms of SASA, radius of gyration and number of hydrogen bond were calculated. Figure 4 depicts the total SASA of each protein, i.e., the biomolecular surface area that is assessable to solvent molecules. Here, decreased values of SASA denotes the shrunken nature [71]. The results of SASA analysis suggest that ZINC02585769 increased the SASA of protein with several large fluctuations in 100 ns simulation. Conversely, ZINC13370550 showed a pattern similar to that of the standard inhibitor sunitinib, indicates that ZINC13370550 was dynamic over the simulation. Comparatively, ZINC36470466 showed lower SASA values over time, allowing us to conclude that the ligand caused tight packing of the protein. Interestingly, ZINC14762520 maintained a steady SASA profile over time; however, some large fluctuations were seen in 100 ns of simulation. On the other hand, the radius of gyration of the proteins, shown in Fig. 5a, revealed that ZINC14762520 produced a higher radius of gyration at the beginning, which was then maintained over the entire simulation time. In comparison to the other ligands, ZINC36470466 showed lower values; these results therefore suggest flexibility of ZINC14762520 as well as tight packing of ZINC36470466, as higher radius of gyration value denotes loose packing of the protein structure, which clearly supports the results calculated from SASA analysis. In addition, sunitinib maintained a radius of gyration profile similar to that of the apo state after 17 ns; no significant changes were seen thereafter. Figure 5b represents the total intra-residue hydrogen bonds formed during the simulations, in which maximum highest picks were observed for apo form and ZINC14762520. On average, there was insignificant variation of hydrogen bonding among the proteins, ranging from 50 to 100 (Table S2, see supplementary file 1). Nevertheless, it is concluded that binding of ligands, including standard inhibitors, reduced intra-residue hydrogen bonds in the protein, therefore increasing flexibility. The degree of flexibility increased by binding of ZINC13370550 and thus the overall dimension of the structure was enhanced, while it was reduced by the binding of ZINC36470466. The total number of H-bonds between the ligand and residues of the ATP binding site of protein for each complex are represented in Fig. 6, where maximum hydrogen bonding was observed for ZINC36470466 (Fig. 6e) during the simulation, while ZINC13370550 (Fig. 6c) showed minimum interaction (Table 2). For a detailed understanding, H-bond occupancy was calculated and represented in Fig. 6f. Here Fig. 6f shows the percentage of H-bond occupancy, where each residue either acted as a donor or as an acceptor.

Fig. 4
figure 4

Ligand-induced thermodynamic behavior of protein during simulations. The plots describe structural changes of protein by means of solvent accessible surface area (SASA). Complexes: Blue apo, orange sunitinib, red ZINC02585769, olive green ZINC13370550, black ZINC14762520, pink ZINC36470466

Full size image
Fig. 5
figure 5

Time evolution of a radius of gyration, and b total number of hydrogen bonds formed within the protein during 100 ns of MD simulations. Complexes: Blue apo, orange sunitinib, red ZINC02585769, olive green ZINC13370550, black ZINC14762520, pink ZINC36470466

Full size image
Fig. 6a–f
figure 6

Total number of hydrogen bonds formed between protein and ligand in complex state during the simulation. a Sunitinib, b ZINC02585769, c ZINC13370550, d ZINC14762520, e ZINC36470466. f Percentage occupancy of hydrogen bond formed between the ligands and active site residues, showing the time evolution of the occupancy of the listed hydrogen bonds, where each residue either acted as donor or acceptor in hydrogen bonding with ligand

Full size image

As shown in Fig. 6f, ZINC14762520 formed several H-bonds with the residues of DFG motif and hinge region, including C919, E885, F1047 and D1046 residues during the simulation, with total occupancy of more than 15%. The major interaction with C919 was mediated by the hydroxyl group present in the 4-hydroxyphenyl part of ZINC14762520, where the oxygen atom acted as acceptor. ZINC36470466 formed H-bonds with D1046 and E885 residues as H-bond donor, occupying more than 50% of the simulation, while less than 20% of H-bonds were occupied as acceptor with C919 residue. Here, hydroxyl groups present at the 3-hydroxybenzyl part of the compound act as donors in the interactions with D1046 residues of DFG motif. The main contribution of ZINC02585769 in the hydrogen bonding interaction was with residue E885, with 121.04% H-bond occupancy. The hydroxyl groups present at the 3,4-dihydroxyphenyl part are major contributors in H-bond formation, acting as donors. In contrast, ZINC13370550 showed maximum H-bond occupancy with T926 and C919 residues as both H-bond donor and acceptor, with interaction maintained by the d-glucopyranose ring and ketone group present in the flavones part, respectively. Sunitinib interacted with residue E917 more than 35% as a donor with an additional H-bond with L840 having a percentage occupancy of around 10%. As an acceptor, sunitinib maintained H-bonds with C919 residues about 34% of total simulation time.

In computer-aided drug design approaches, Poisson-Boltzmann surface area (PBSA) is one of most attractive solvation systems that is applied widely to calculate binding energy of protein–ligand complexes [51, 72]. Hence, we subjected each protein-ligand complex to a MM-PBSA binding energy calculation to see how structural changes influence ligand binding. With this aim, we calculated the binding energy of every snapshot produced from MD simulations. The results are plotted in Fig. 7a, where more positive energies indicated better binding, according to the theory of nuclear physics [73]. According to the laws of physics, the energy required to disassemble a whole into separate parts is usually positive [74]. On average, ZINC14762520 obtained maximum energy of binding of 166.12 ± 6.5 kJ mol−1, while ZINC02585769 showed the least energy of 19.45 ± 1.4 kJ mol−1. The reference inhibitor exhibited a binding energy of 115.43 ± 1.2 kJ mol−1 By contrast, the ZINC13370550 and ZINC36470466 complexes showed positive binding energy values of 49.52 ± 5.2 kJ mol−1and 60.92 ± 4.7 kJ mol−1, respectively.

Fig. 7
figure 7

a Binding free energy (in kJ mol−1) of each snapshot was calculated by molecular mechanics Poisson-Boltzmann surface area (MM-PBSA) analysis, representing the change in binding stability of each complex during simulations; positive values indicate better binding. b Time evolution of root-mean-square deviations (RMSDs) of ligand for each docked complex. Complexes: Blue apo, orange sunitinib, red ZINC02585769, olive green ZINC13370550, black ZINC14762520, pink ZINC36470466

Full size image

Although ZINC02585769 produced higher binding energy at initial stages of simulation, the binding energy was seen to decrease afterward. ZINC36470466 and ZINC14762520 showed similar patterns of binding energy profiles; however, ZINC36470466 obtained lower binding energy after 50 ns of simulation. As ligand flexibility is involved in the formation and breakdown of new bonds in the active site, we analyzed ligand RMSD to gain more detailed understanding. As shown Fig. 7b, the conformation of the sunitinib fluctuated up to 2.4 Å, and several large decreases in fluctuation were observed one by one over time. Despite several fluctuations, sunitinib maintained binding energy above 40 kJ mol−1 in total time of MD simulation. Likewise, over the course of simulation, ZINC14762520 supported stable conformations with RMSD up to 0.7 Å, although several oscillations were observed, which therefore caused variations in binding energy. With regard to ZINC13370550, the ligand showed conformational shifting at 19 ns, which thus produced higher binding energy at that time; however, several large shifting occurred afterward which indicates greater fluctuation in binding energy profile. As a result, these conformational shifts caused weaker binding of the ligand with the active site residues of the protein, therefore resulting in low MM-PBSA binding free energy, as can be observed in Fig. 7a. On the other hand, ZINC02585769 showed a rigid conformation over time, thus showing that flexibility of the ligand plays an important role in the formation of a stable protein–ligand complex.

Overall, MD trajectory analysis infers that ZINC14762520 and ZINC36470466 are the most potent candidates among the herbal secondary in vivo metabolites to inhibit VEGFR-2 kinase. Both these compounds achieved highest average binding free energies in MM-PBSA analysis among the top virtual hits, and the binding energy of ZINC14762520 was greater than the reference inhibitor sunitinib. They also maintained maximum non-covalent interactions in 100 ns with residues of the three most important ligand binding regions in VEGFR-2 kinase, including G-loop, DFG motif and hinge regions [75]. In all kinase domains, G-loop and hinge regions serve as important components for ligand binding, where the G-loop interacted with the purine ring of adenine, and the hinge region acts as the ribose binding pocket [76]. Interestingly, both compounds ZINC14762520 and ZINC36470466 interacted with the residues involved in the hinge region and the DGF motif. Comparatively, ZINC36470466 sustained maximum interactions with hinge regions by its benzyl rings, although ZINC14762520 contains a hydroxyl pharmacophore in the 4-hydroxyphenyl portion that stabilized hinge binding by hydrogen bonding more than 35% in total simulation (Fig. 3f). Furthermore, docking analysis showed that ZINC36470466 formed hydrophobic interactions through its aromatic ring with L840 and V848 residues of the G-loop region. Hence, we calculated the atomic distances between the interacting ring of the ligands and the beta carbon of L840 and V848 residues, respectively. As shown in Fig. S4 (see supplementary file 1), ZINC36470466 maintained distances less than 5.5 Å in 100 ns simulation with these hydrophobic residues, despite exhibiting some fluctuation. Furthermore, L840 maintained maximum intermolecular distance with the ligand from 4 to 5.5 Å in the total simulation, which indicates that L840 of the G-loop contributed stronger hydrophobic interaction than V848. Taken together, the above analysis suggests strong stability and VEGFR-2 kinase inhibiton with both ZINC14762520 and ZINC36470466.

Conclusion

In this in silico experiment, we tried to identify potent inhibitors of VEGFR-2 from a database of in vivo metabolites of herbal ingredients, using Glide XP docking and ADME/T analysis to identify the six best candidates. Among them, the top four virtual hits were specified by MM-GBSA calculations. And finally, ZINC14762520 and ZINC36470466 have been selected as the best virtual hits by MM-PBSA and MD simulations. Molecular docking analysis revealed the detailed binding mode of the selected virtual hits with VEGFR-2. According to MD simulation analysis, these compounds energetically formed stable complexes with VEGFR-2 kinase owing to different conformational changes. In addition, all identified compounds were predicted to be orally active and satisfied a drug likeness profile. Compared to standard inhibitors, they are also predicted to be less susceptible to cardiac toxicity. The results of this study therefore offer a suitable starting point for further development as well as detailed in vitro and in vivo analyses.