Prediction of new Hsp90 inhibitors based on 3,4-isoxazolediamide scaffold using QSAR study, molecular docking and molecular dynamic simulation



Heat shock protein90 (Hsp90) are overexpressed in tumor cells, so the inhibition of the Hsp90 ATPase activity would be a significantly effective strategy in cancer therapy.


In the current study, 3,4-isoxazolediamide derivatives were suggested as an Hsp90 inhibitor for anti-cancer therapy. Multiple linear regression (MLR) and genetic algorithm of partial least square (GA-PLS) methods were performed to build models to predict the inhibitory activity of Hsp90. The leave-one out (LOO) cross-validation and Y-randomization tests were performed to models’ validation. The new ligands were monitored by applicability domain. Molecular docking studies were also conducted to evaluate the mode of interaction of these compounds with Hsp90. Identification of the likely pathways into the active site pocket and the involved residues were performed by CAVAER 3.0.1 software. According to QSAR models and docking analysis, three new compounds were predicted. 50 ns molecular dynamic simulation was performed for the strongest synthesized compound and the best predicted compound in terms of binding energy and interactions between ligand and protein.


The made models showed the significance of size, shape, symmetry, and branching of molecules in inhibitory activities of Hsp90. Docking studies indicated that two hydroxyl groups in the resorcinol ring were important in interacting with Asp93 and the orientation of these groups was related to substitution of different R1 groups. Comparing of molecular dynamic simulation (MDs) results shows that new compound perched in active site with lower binding energy than the best synthesized compound.


The QSAR and docking analyses shown to be beneficial tools in the proposal of anti-cancer activities and a leader to the synthesis of new Hsp90 inhibitors based 3,4-isoxazolediamide. The MDs confirmed that predicted ligand is steady in the Hsp90 active sites.

Graphical Abstract


Heat shock proteins (Hsps) play a critical role in maturation and stabilization of proteins in the cell [1]. One of the important Heat shock proteins in cell is Hsp90. The Hsp90 molecular chaperone contributes to folding of more than 200 proteins (client proteins) and it is necessary for adjusting the balance between the synthesis and degradation of many proteins in the cell [2, 3]. The homo dimer of Hsp90 possess three main domains: the N-terminal domain that contains the nucleotide-binding pocket, the middle region that is involved in binding of client proteins and the C-terminal domain that is the dimerization site [4]. The Hsp90 function is reliant on its ability to bind and hydrolyze ATP at the N-terminal domain. First, client protein and co-chaperones bind to Hsp90 in the open condition of protein and then ATP binds to N-terminal and Hsp90 will be closed. Finally, ATP is hydrolyzed, the complex is changed and client protein is folded (Fig. 1) [5].

Fig. 1

ATPase cycle in Hsp90. The first, in open state of protein, co-chaperones and client protein bind to C-terminal and middle domain, respectively. ATP bind to N-terminal and Hsp90 is closed. Then ATP is hydrolyzed and the complex changed. Finally, client protein is folded

In tumor cells, Hsp90 is overexpressed and causes the uncontrolled proliferation of transformed cells [6], so inhibition of the Hsp90 ATPase activity can be a significantly effective strategy in cancer therapy [7]. Hsp90 inhibitors are classified into several categories containing natural inhibitors (geldanamycin, GM (1), and radicicol (2)), reclaimed analogues of GM (17-AAG (3) and 17-DMAG (4), synthetic inhibitors (purine (PU3 (5)), pyrazole (6), indazole (7), aminoquinolines (SID: 24724290 (8)) and isoxazole (9) that are shown in Fig. 2 [8,9,10,11]. Some of these inhibitors, such as the reclaimed analogue of geldanamycin (17-AAG), carbazol-4-one benzamide derivative (SNX-5422) and isoxazole derivative (NVP-AUY922, currently known as Luminespib), have been assessed in humans (Fig. 3) [1]. Among the different azaheterocyclic ring systems, the isoxazole scaffold is one of the most promising heterocyclic systems [12]. Baruchello and coworkers in 2011 synthesized novel 3, 4-isoxazolediamides as potent inhibitors of Hsp90 [8].

Fig. 2

Several categories of Hsp90 inhibitors. Natural inhibitors (geldanamycin, GM, (1), radicicol (2)), reclaimed analogues of GM (17-AAG (3) and 17-DMAG (4)), synthetic inhibitors (purine (5, PU3), pyrazole (6), indazole (7), aminoquinolines (8, SID: 24724290) and isoxazole (9))

Fig. 3

The structure of three clinically Hsp90 inhibitors evaluated in human. Derivative of geldanamycin (17-AAG), carbazol-4-one benzamide (SNX-5422) and isoxazole (NVP-AUY922)

Forasmuch as obtaining a new inhibitor requires much time and capital, any tool that can help to precipitate the drug development processes would be very noteworthy [13]. The advanced computational techniques are highly useful strategies to conduct rapid and inexpensive investigations on large databases and obtaining new inhibitors [14].

Molecular docking, molecular dynamic simulation and quantitative structure activity relationships (QSARs) are helpful computational methods for drug design and activity prediction [15, 16]. In molecular docking and molecular dynamic simulation, the 3D structure of the receptor will be available and receptor-ligand interactions play an important role, so this drug design is called structure-based drug design [17]. Docking is a method which proposes the favored orientation and energy of one ligand when bound in the active site to build a stable complex. To investigate interactions between ligand and protein, molecular dynamic simulation will be performed.

QSAR is referred to as ligand-based drug design because it is performed based on the knowledge about ligands [18]. QSAR models are mathematical equations that create relationships between chemical structures and their biological activities without considering a receptor.

In the present study, we performed a QSAR study for modeling the inhibitory effects of 50 synthesized 3,4-isoxazolediamides [8]. Multiple linear regression (MLR) and genetic algorithm of partial least square (GA-PLS) methods were used for modeling the relationship between pIC50 and their structural descriptors. Molecular docking was performed for 25 compounds and the results were compared with experimental data. According to QSAR models and molecular docking analysis three compounds were predicted. Molecular dynamic simulation method was chosen to compare the best synthesized and the best predicted compound in terms of binding energy and interactions between protein and ligand.


Activity data and generation of descriptors

In the current study, the Hsp90 inhibitory activity of 50 compounds that were synthesized and evaluated by Baruchello and coworkers, were used as the biological data [8]. To make the mistake less, the compounds were chosen from an article with the same assay. Reported compounds were chosen from isoxazole derivatives. The lead of these compounds was NVP-AUY922 (currently known as Luminespib) that was entered into clinical trials and many of its derivatives were synthesized up to now. These compounds are shown in Table 1.

Table 1 Chemical structures of 3, 4-isoxazolediamides derivatives as Hsp90 inhibitors

Geometry optimization was performed by two methods: MM+ molecular mechanic force field and the semi-empirical PM3 by HyperChem 8.0 software. Both optimizations were carried out by the Polak–Ribiere algorithm until the root mean square gradient reached 0.01 kcal/Å mol. Chemical descriptors were also generated by HyperChem software.

HyperChem outputs were conveyed to DRAGON package to compute four classes of descriptors: 0D, 1D, 2D and 3D. Different quantum chemical descriptors were obtained by Gaussian98 program [19]. All of the calculated descriptors are shown in Table 2.

Table 2 Brief of used descriptors in this study

Data processing and model building

All of the calculated descriptors were used to generate a 50 × 1126 data matrix; the number on its rows is illustrative of molecule numbers and the numbers on columns accounts for descriptors. The columns which had constant and near constant amounts were deleted from this data matrix. Since as the models were disrupted by collinear variables, collinear descriptors had to be found and eliminated and then descriptors’ correlations with one another and with activity data were checked. In couples with collinearities higher than 0.9, the highest correlation with the activity was kept and the rest were eliminated. The number of total descriptors for each molecule attained 352.

The data set (50 compounds) was divaricated into the calibration set and validation set. Validation subset was prepared from 20% of the total data (here, 9 biological activity data). The model was made by MLR analysis, with the stepwise regression SPSS (Version 12.0) software.

The combined data splitting feature selection (CDFS) approach was utilized because the data splitting has a significant effect on the terminal selected model [15]. In the CDFS approach, numerous subdivisions of the calibration and validation set were created (5 times). In each case, the best model was elected. The chosen models were validated by leave-one out (LOO) cross-validation method. According to 5 models descriptors, general model was created and Y-randomization test was performed to check their predictability.

Molecular docking

Molecular docking of 3,4-isoxazolediamide derivatives (25 compounds) as Hsp90 inhibitors was perused by AutoDock 4.2 program to detect their binding site, the best direction and the binding energy [20]. Between the experimental X-ray structures of Hsp90, the crystallography structure with a PDB entry code of 3OWD was chosen [21].

For protein preparation, the co-crystallized ligand and water molecules, except the water molecules that were significant in interaction between the ligand and protein, were deleted. By AutoDockTools, all missing hydrogens were added. The Kollman atom charges were calculated, non-polar hydrogens merged and the file saved as pdbqt. A grid box was created with a grid point spacing of 0.375 Å and 90 × 90 × 90 points, which included not only the active site of the protein but also significant regions of the surrounding surface. Before calculating the grid maps by AutoGrid 4.2 [22], the parameters of the water molecules were added to AD4-bound and AD4-parameter files.

After the ligands were prepared, the 3D structures of all compounds were dragged in Marvin Sketch Ver. 5.7, ChemAxon [23]. The partial charges of atoms were computed by the Gasteiger–Marsili procedure and non-polar hydrogens of the compounds were merged [24].

The Lamarckian genetic algorithm approach was elected for the global optimum binding position. Docking computing parameters were determined as the following: number of Lamarckian job = 50, initial population = 150, maximum number of energy evaluation = 25 × 105, and the default values of other parameters were kept unchanged. The docking parameter file (.dpf) was created. The docking procedure was done by AutoDock 4.2 and the .dlg file was generated. All of the runs were ranked by the maximum number of clusters and the lowest binding energy and were analyzed to find the best conformation of the ligand with key residues in the active site of the protein by Accelrys Discovery Studio 2.5 [25] and PyMOL software [26].

Caving active site tunnels

To monitor likely pathways into the active site pocket of the protein and also recognize the involved residues, CAVAER 3.0.1 software was employed [27]. Likely entrances were searched by assigning maximum probe radius to 0.9Ǻ, shell depth to 4, and clustering threshold to 3.5Ǻ.

Molecular dynamic simulation

Molecular dynamic simulation was carried out with the GROMACS 5.0.5 package [28]. The topology parameters of the best predicted ligand in terms of the energy and interaction between ligand and protein were created by the PRODRG web server [29]. The generated charges by this server were corrected by Gaussian98 program. The pKa for residues of protein were obtained by the PROPKA 3.1 web server to determined which residue was more possible to embrace nonstandard ionization states [28]. The key crystallographic water molecules in the active site were kept [25]. The GROMOS96 54a7 force field and the simple point charges (SPC) water model were used to create protein topology parameters. The complex of ligand and protein was dunked in dodecahedron box with a minimum distance of 1 nm between the protein surface and box border, containing of about 8250 solvent molecules. By displacing solvent water molecules with 4 Na + was neutralized the system net charge. The energy minimization was done to release spatial clashes of the complex in two steps. First, only water molecules were minimized by using 10,000 steepest descents steps while the other atoms were hold fixed at their initial configuration. After that, the entire system was minimized. To equilibrate the system at a constant temperature of 300 K was performed NVT step by a 500 ps MD run. After the stabilization of temperature by the V-Rescale algorithm, an NPT ensemble was done with time duration of 1 ns. This was followed by MD production run at 1 bar pressure and 100, 200 and 300 K temperatures for 1, 2 and 50 ns, respectively. Long-range electrostatic interactions were computed with the Particle Mesh Ewald (PME) method. The linear constraint (LINCS) algorithm was used for covalent bond constraints. Structure visualization was carried out by VMD 1.8.6 and PyMOL.

Result and discussion

Multiple linear regression analysis (MLR)

Stepwise regression method was used to select the most appropriate set of descriptors for each type of the split data. Models were selected; of course, some of them could be over-fitted. Another method which could be utilized to select the model or the most suitable correlation equation was the cross-validation method. In this case, the obtained models were evaluated for over-fitting by leave-one-out cross-validation (LOO) method and then were consistently graded for cross-validation (Q2 LOO) by the square correlation coefficient. Eventually, one model with equilibrium between the highest R2 c and Q2 LOO was chosen for future analysis.

MLR method was performed five times by distinct split data. In each case, one model was offered. The best five models are reported in Table 3. The models demonstrate high statistical qualities. All of the models have Q2 LOO greater than 0.5; hence, the expected models indicated suitable results for the prediction set. The values of prediction correlation coefficient (R2 p) for the five final models are documented in Table 3.

Table 3 The best five models were selected for future analysis

The total number of descriptors that existed in all five models was 25. These descriptors are briefly defined in Table 4. Among these, X5A and 3D–MoRSE descriptors have been repeated in four models. This means that connectivity indices (X5A) and 3D–MoRSE descriptors have main effect on Hsp90 inhibitors which are based on 3, 4-isoxazolediamide scaffold. Some of the descriptors such as Du, BELe1, RDF140m, Gu, ISH, Ku, P2m, and TIE were observed only in one model, and have a lower effect on Hsp90 inhibitors.

Table 4 Brief description of the descriptors in five models

To generate a general model for split data, all of the 25 descriptors were applied and MLR analysis was conducted with the stepwise variable selections. The resulted model is reported in MLR equation:

$$ {\mathrm{pIC}}_{50}=16.977\left(\pm 1.966\right)-176.806\left(\pm 28.236\right)\ \mathrm{X}5\mathrm{A}-4.366\left(\pm 1.046\right)\ \mathrm{HATS}4\mathrm{u}-1.951\left(\pm 0.549\right)\ \mathrm{Mor}26\mathrm{p}+0.001\left(\pm 0.00\right)\ \mathrm{TIE}+0.227\left(\pm 0.069\right)\ \mathrm{dipole}\ \mathrm{z}+0.863\left(\pm 0.368\right)\ \mathrm{Mor}26\mathrm{e} $$
$$ \left(\mathrm{N}=41,{{\mathrm{R}}^2}_{\mathrm{c}}=0.771,\mathrm{S}.\mathrm{E}=0.267,{{\mathrm{Q}}^2}_{\mathrm{LOO}}=0.637,{\mathrm{R}\mathrm{MS}}_{\mathrm{CV}}=0.297\right) $$

Where N demonstrates the number of molecules used in the calibration set. R2 c and Q2 LOO are respectively the squared correlation coefficient for calibration and cross-validation. In addition, S.E is standard error of calibration and RMSCV is root mean square error of cross-validation.

In this equation, X5A, HATS4u and Mor26p have a minus, which indicates that the pIC50 is inversely related to these descriptors. Among these, X5A has a main effect on 3, 4-isoxazolediamide Hsp90 inhibitors. Connectivity indices, such as X5A, are among topological indices that are numerical quantifiers of molecular topology and an H-depleted molecular graph. They involve one or more structural features of the molecule such as the size, shape, symmetry, and branching and can also codify chemical information about atom type and bond multiplicity. 3D–MoRSE descriptors, such as Mor26p, illustrate three-dimensional coordination of the different atoms in the molecule. Geometry, topology, and atom-weights assembly (GETAWAY) descriptors, such as HATS4u, have been proposed as chemical structure descriptors.

TIE and dipole z display positive signs, which indicates the pIC50 is straight related to these descriptors. Topographic indices such as TIE establish a special subset of geometrical descriptors, being computed on the graph illustration of molecules but using the geometric distances between atoms instead of the topological distances [30].

The general model has a Q2 LOO equal to 0.710; hence, the predicted model can construct over 71% of variances in the inhibitory activity. The proposed values of pIC50, which were calculated for all the molecules by the MLR equation, along with the experimental pIC50 are listed in Table 5 and the predicted values of pIC50 are plotted against the experimental values in Fig. 4a.

Table 5 Experimental pIC50 and MLR and GA-PLS predicted pIC50
Fig. 4

a Plot of predicted pIC50 versus the experimental values for MLR model, b GA-PLS model

The robustness of the general model was evaluated by the Y-randomization test and 10 models were created. In a suitable MLR model, the values of R2 and Q2 LOO are lower than the original MLR model. The lower R2 and Q2 LOO values are shown in Table 6.

Table 6 R2 and Q2 LOO values after ten Y-randomization tests in MLR and GA-PLS

Genetic algorithm partial least squares (GA-PLS)

In addition to MLR method, the stepwise regression, a variable selection method, such as genetic algorithm, was performed. In genetic algorithm, if its related descriptor is comprised, a gene takes a value of 1 in the subset; otherwise, it receipts a value of zero. The number of genes is corresponding to the number of descriptors.

In GA-PLS study, the total number of descriptors for each molecule attained 117 and the highest fitness was selected as the proposed model. The intended model was tested by leave-n-out cross-validation [31]. A leave-3-out cross-validation was generated and Q2 LTO value of 0.834 could be established.

The eventuated model with higher cross-validation statistics is shown in the following equation and the predicted values of pIC50 are reported in Table 5 and plotted against the experimental values in Fig. 4b.

$$ \mathrm{pIC}50=9.018\left(\pm 2.799\right)-114.521\left(\pm 30.435\right)\ \mathrm{X}5\mathrm{A}-1.406\left(\pm 0.266\right)\ \mathrm{G}\mathrm{ATS}4\mathrm{e}-1.586\left(\pm 0.991\right)\ \mathrm{E}3\mathrm{u}-2.615\left(\pm 0.649\right)\ \mathrm{MATS}7\mathrm{e}+38.705\left(\pm 8.531\right)\ \mathrm{G}1\mathrm{u}+0.048\left(\pm 0.014\right)\ \mathrm{RDF}075\mathrm{m} $$
$$ \left(\mathrm{N}=41,{\mathrm{R}}^2\mathrm{c}=0.755,\mathrm{S}.\mathrm{E}=0.249,{{\mathrm{Q}}^2}_{\mathrm{LOO}}=0.653,\mathrm{RMSCV}=0.266\right) $$

In this equation, X5A descriptor, like MLR equation, has a minus and also has a main effect on 3, 4-isoxazolediamide Hsp90 inhibitors. GATS4e, E3u, and MATS7e also have minuses. Unlike these, G1u and RDF075m (WHIM and the radial distribution function (RDF) descriptors) have a positive efficacy on GA-PLS.

The created model in GA-PLS was evaluated by Y-randomization test. The R2 and Q2 LOO values of the ten created models were lower than the original GA-PLS model, as shown in Table 6.

Applicability domain of the model

A QSAR model is applied to monitor new compounds when its domain of application has been determined [29]. The divination may be supposed valid for only those compounds which fall into this domain [15]. Therefore standardized residuals of the activity were calculated and plotted against leverage values (h). The value of leverage was computed for every compound. Values are evermore between 0 and 1. A value of 1 indicates very poor prediction, and a value of 0 is indicative of perfect prediction and is not usually available. The lower the value is, the higher the assurance in the prediction. Warning leverage (h*) is another standard for description of the results and is commonly stabilized at 3 (k + 1)/n, where k is the number of model parameters and n is the number of calibration set [15]. Obtained leverage for calibration set is used to define the compounds which affect the model and, in terms of validation set, used to determine the applicability domain of the model. William’s plot for the advanced models in MLR and GA-PLS is revealed in Fig. 5.

Fig. 5

a William’s plot of generated MLR model, b GA-PLS model

Rejoinder outliers are compounds that have standard residual points higher than ±2.5 and ±3.0 standard deviation units for MLR and GA-PLS respectively and also a leverage value higher than the warning leverage, which is 0.42. Figure 5 shows that all the considered molecules in calibration and validation set lie with a high degree of assurance in the application domain of the developed models in both methods.

Molecular docking

ATP is Hsp90 substrate and its active site includes mixed hydrophobic, polar and charged amino acids. These residues include Leu48, Asn51, Asp54, Ala55, Lys58, Ile91, Asp93, Ile96, Gly97, Met98, Asn106, Leu107, Lys112, Gly135, Phe138, Val150, Thr184, and Val186. While the pocket becomes increasingly hydrophobic toward the bottom, one charged residue and one polar residue are retained as Asp93 and Thr184 respectively [32]. The adenine ring in ATP sits at the bottom of the pocket and its N6 group form direct hydrogen bonds with Asp93 (Asp79 in yeast Hsp83), and water-mediated hydrogen bonds with the Ser52 and Leu48. Also, Glu47 (Glu33 in yeast Hsp83) gets involved in Hsp90 ATPase activity [33, 34].

A crystallography structure of 3OWD was chosen for the structure-based drug design. A validation of docking procedure was done by the extraction of the ligand from X-ray complex and redocking it. A .dlg file was created. A cluster root-mean-square deviation (RMSD) at less than 2 Å whose initial coordinates of the ligand were used as the reference structure was observed. The obtained results show that the docked ligand was located in the active site of Hsp90 (Fig. 6).

Fig. 6

a The position of ligand in X-ray crystallography, b The position of ligand after redocking

As previously mentioned, Asp93 is one of the important residues in Hsp90 ATPase. The pathway into Asp93 is a large cavity, and in its crater Asn51, Ala55, Lys58, Gly97, Met98, Asp102, Asn106 and Phe138 lie. Asp93 and Thr184 stay on the bottom of the pocket (Fig. 7a). Other pathways into Asp93 and the involved residues were obtained by CAVAER 3.0.1 software [27]. Likely pathways were computed by assigning maximum probe radius to 0.9 Ǻ, shell depth to 4, and clustering threshold to 3.5 Ǻ. Four tunnels were obtained, which are depicted in Fig. 7b. As can be seen, all of the tunnels stay at the back of the cavity.All residues in the depth of the tunnel included Ile49, Ser50, Ser52, Ser53, His77, Ile78, Leu80, Ile91, Val92, Asp93, Val207, His210, Gln212, Phe213, Ile214, Tyr216, Pro217, Ile218, Thr219 and Leu220.

Fig. 7

a Main cavity in Hsp90, b Four entrance tunnels in back of main cavity

After the validation of the docking protocol and finding entrance tunnels, the 3D structures of 3,4-isoxazolediamide derivatives (25 compounds) were docked into crystallography structure 3OWD. In .dlg files, the structures were ranked by binding energy and clusters. All of the compounds were perched into the cavity of the active site (Table 7). The compounds were divided into three groups.

Table 7 Interactions between the docked 3, 4-isoxazolediamide derivatives and Hsp90 binding site residues

Group A (compounds 15–24)

In this Group, the meta-position of resorcinol ring was substituted by a large chloro group. So, when R1 group was small, for example like the CH3 group, hydroxyl groups in resorcinol ring could interact well with Asp93. On the other hand, large groups in R1 such as 1,2,3-trimethoxybenzene groups, due to steric hindrance, changed steric orientation of hydroxyl groups in resorcinol and hydroxyl groups could interact with Asp93 by a water-bridge hydrogen bond. According to what was mentioned above, compound 20 was the best inhibitor in this group. Compound 18 had a 1,2,3-trimethoxybenzene group, but trimethoxy groups created a better hydrogen bond with Lys112 than methoxy group.

Group B (compounds 26–73)

Meta position of resorcinol ring was occupied by an isopropyl in this group. According to the biological assay, the lowest IC50 was related to compound 69 and it was the strongest compound that docking analysis revealed. The ortho-hydroxyl group of resorcinol ring with Asp93, para-hydroxyl with Ser52 by a water-bridge, and nitrogen atom of isoxazole with Gly97 was created the hydrogen bonds. Steric orientation in compound 58 was observed to be similar to that of 69. But, 69 was stronger than 58, because sulfur atom of R1 group in 69 created a hydrogen bond with Phe138 and Asn51 by a water-bridge. In compound 46, para-hydroxyl made a hydrogen bond with Asp93 and O atom in methoxy group could create H-bond with Lys58. The steric orientation in cyclohexane made compound 28, and 30 could not directly create a hydrogen bond with Asp93. The connected R1 group to carbonyl was isoxazole in 73 and 59 but 73 was a stronger inhibitor than 59. The connected benzene group to this isoxazole changed the steric orientation in resorcinol and could not directly make a hydrogen bond with Asp93. Unlike Group A, in this group the compounds with smaller R1substituent were not potent inhibitors.

Group C (compounds 81a-81c)

In this group, on meta-position of resorcinol ring was placed a large chloro group and R1 was connected to NH. 81a was stronger than both 81b and 81c. Thiophene group was better than either methoxybenzen or cyclohexane. The same as Group A, the compounds with small R1 groups were stronger inhibitors.

Predicted ligands and molecular dynamic simulation

Three novel compounds (Fig. 8) were predicted with the modification of compounds based on results of QSAR model and molecular docking.

Fig. 8

The predicted novel inhibitors based on QSAR model and docking

Docking study was performed on the predicted compounds. In these compounds, similar to previously evaluated compound, the hydroxyl groups of resorcinol ring interacted with Asp93 (Fig. 9). The binding energy of compound II was the lowest (−8.20 kcal/mol).

Fig. 9

Orientation and main interaction between three predicted compounds and Hsp90

To ensure the stability of the new ligand II in the active site of protein, the MD simulation was carried out and its interaction modes with compound 69 in Hsp90 active site was also compared. The best conformation of docking was chosen to perform 50 ns MD simulation. The time-dependent behavior of MD trajectories were analyzed including root mean square deviation (RMSD) for all backbone atoms and ligands, average fluctuations of the residues (RMSF) and gyration radius (Rg) for all backbone atoms.

The RMSD of backbone atoms was figured to assessment the conformational stability of the protein during the simulation. The RMSD of backbone atoms between complexes of Hsp90-ligand 69 and Hsp90-predicted ligand II was measured with a 10 ps time interval. As illustrated in Fig. 10a, the RMSD profile in the two complexes was almost the same in the first 10 ns. Variations of RMSD were not very considerable, which denote the stability of both complexes. The ligand RMSD profile in Fig. 10b shows that the ligand 69 after 3 ns and predicted ligand II after 8 ns become stable. The RMSD variations were almost 0.1 nm in ligand 69 and predicted ligand II, it can be observed that both of the ligands were also fit in the active site and stabilized. The protein compactness was evaluated by the gyration radius (Rg). The Rg plot is shown in Fig. 11. In the first 27 ns and the last 10 ns, the Rg values of ligand 69 and the predicted ligand II were superimposed and the conjunction of both complexes was conserved during the simulation.

Fig. 10

The RMSD profile. a Hsp90 backbone in complex with compound 69 (violet), predicted ligand II (green). b Compound 69 (violet), predicted ligand II (green) as a function of simulation time

Fig. 11

The gyration radius plot of backbone. Compound 69 (violet), predicted ligand II (green)

The variations of protein flexibility were identified by the root mean square fluctuation (RMSF) of backbone residues. As shown in Fig. 12, in residues 45–61 and 89–107, the fluctuation of predicted ligand II was lower than compound 69 which disclose that predicted ligand II was more stable than compound 69 in these parts. This can be attributed to the attendance of these residues in the active site of protein. In residues 61–70 the fluctuation was high, which indicates that these parts of the protein were more unstable than other parts during the MD simulation, especially in the predicted ligand II.

Fig. 12

The RMSF plot. Hsp90-compound 69 (violet), Hsp90-predicted ligand II (green)

To discover the conformations and interactions of the predicted ligand II and compound 69 during simulation, 3D plots of the different times of simulation were shown (Fig. 13). Figure 13 shows the comparison between the compound 69 and predicted ligand II in the active site of Hsp90 during 0, 25 and 50 ns of the simulation (a, b columns). In all simulation times, the resorcinol hydroxyl group of both compounds formed hydrogen bond with Asp93. On other side of both compounds perched toward out of pocket and was not involved in interactions between ligand and protein. Since the beginning until the half of simulation (0–25 ns), the hydrogen bond was built with crystallography water molecules in both compounds. In compound 69, at the end of simulation (50 ns), the N atom of isoxazole ring and the resorcinol hydroxyl group made hydrogen bonds with Asp93 but the crystallography water molecules was not involved in interactions between ligand and protein. Adding amine group to benzothiophene moiety, conformation of predicted ligand changed and the interaction between ligand and the crystallography water molecule was seen with 1.88 Å distance. According to the MD simulation analysis and also the evaluation of the interactions between compound 69 and predicted ligand II with Hsp90, adding amine group was effective in improving of binding energy of ligand and protein.

Fig. 13

3D plots of the interaction between two ligands and Hsp90 at different times during the MD simulation. Column a The interaction of compound 69 with Hsp90. Column b the interaction of the predicted ligand II with Hsp90


In current study, quantitative relationships between the molecular structure and Hsp90 inhibitory activity of 3, 4-isoxazolediamide derivative were verified by MLR and GA-PLS. In MLR, the connectivity index (X5A) has a significant effect on the inhibitory activity. This method made suitable results for the prediction set. In addition to MLR, genetic algorithm method was used which illustrated much more promising results in comparison with stepwise regression. It can be concluded from the two methods that structural features of the molecule such as size, shape, symmetry, and branching are particularly important in design of new Hsp90 inhibitors based on 3,4-isoxazolediamide scaffold. The molecular docking simulation was also done for 25 compounds in three groups (A, B and C).

Based on QSAR models and molecular docking analysis were predicted three novel compounds. The ligand II was chosen in term of binding energy and steric orientation. Molecular dynamic simulation was done and analyses such as RMSD, RMSF and Rg, displayed that proposed ligand II, is stable in the Hsp90 active sites.



Adenosine triphosphate


Combined data splitting-features election


Genetic algorithm partial least squares


Heat shock protein90


Molecular dynamic simulation


Multiple linear regression

Q2 LOO :

Square correlation coefficient for leave-one-out cross-validation


Quantitative structure activity relationship


Calibration correlation coefficient

R2 p :

Prediction correlation coefficient


Gyration radius


Root mean square error of cross-validation


Root mean square fluctuation


Root-mean-square deviation


Standard error of calibration


  1. 1.

    Ballante F, Caroli A, Wickersham RB III, Ragno R. Hsp90 inhibitors, part 1: definition of 3-D QSAutogrid/R models as a tool for virtual screening. J Chem Inf Model. 2014;54(3):956–69.

    CAS  Article  PubMed  Google Scholar 

  2. 2.

    Liu J, Wang F, Ma Z, Wang X, Wang Y. Structural determination of three different series of compounds as Hsp90 inhibitors using 3D-QSAR modeling, molecular docking and molecular dynamics methods. Int J Mol Sci. 2011;12(2):946–70.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  3. 3.

    Verma S, Singh A, Mishra A. Dual inhibition of chaperoning process by taxifolin: molecular dynamics simulation study. J Mol Graph Model. 2012;37:27–38.

    CAS  Article  PubMed  Google Scholar 

  4. 4.

    Simunovic M, Voth GA. Molecular and thermodynamic insights into the conformational transitions of Hsp90. Biophys J. 2012;103(2):284–92.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  5. 5.

    McLaughlin SH, Sobott F, Z-p Y, Zhang W, Nielsen PR, Grossmann JG, et al. The co-chaperone p23 arrests the Hsp90 ATPase cycle to trap client proteins. J Mol Biol. 2006;356(3):746–58.

    CAS  Article  PubMed  Google Scholar 

  6. 6.

    Galam L, Hadden MK, Ma Z, Ye Q-Z, Yun B-G, Blagg BS, et al. High-throughput assay for the identification of Hsp90 inhibitors based on Hsp90-dependent refolding of firefly luciferase. Bioorg Med Chem. 2007;15(5):1939–46.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  7. 7.

    Jensen MR, Schoepfer J, Radimerski T, Massey A, Guy CT, Brueggen J, et al. NVP-AUY922: a small molecule HSP90 inhibitor with potent antitumor activity in preclinical breast cancer models. Breast Cancer Res. 2008;10(2):R33.

    Article  PubMed  PubMed Central  Google Scholar 

  8. 8.

    Baruchello R, Simoni D, Grisolia G, Barbato G, Marchetti P, Rondanin R, et al. Novel 3, 4-isoxazolediamides as potent inhibitors of chaperone heat shock protein 90. J Med Chem. 2011;54(24):8592–604.

    CAS  Article  PubMed  Google Scholar 

  9. 9.

    Ganesh T, Min J, Thepchatri P, Du Y, Li L, Lewis I, et al. Discovery of aminoquinolines as a new class of potent inhibitors of heat shock protein 90 (Hsp90): synthesis, biology, and molecular modeling. Bioorg Med Chem. 2008;16(14):6903–10.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  10. 10.

    O’Boyle NM, Knox AJ, Price TT, Williams DC, Zisterer DM, Lloyd DG, et al. Lead identification of β-lactam and related imine inhibitors of the molecular chaperone heat shock protein 90. Bioorg Med Chem. 2011;19(20):6055–68.

    Article  PubMed  Google Scholar 

  11. 11.

    Taldone T, Patel HJ, Bolaender A, Patel MR, Chiosis G. Protein chaperones: a composition of matter review (2008-2013). Expert opinion on therapeutic patents. 2014;24(5):501–18.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  12. 12.

    Tzanetou E, Liekens S, Kasiotis KM, Melagraki G, Afantitis A, Fokialakis N, et al. Antiproliferative novel isoxazoles: modeling, virtual screening, synthesis, and bioactivity evaluation. Eur J Med Chem. 2014;81:139–49.

    CAS  Article  PubMed  Google Scholar 

  13. 13.

    FitzGerald GA. Anticipating change in drug development: the emerging era of translational medicine and therapeutics. Nat Rev Drug Discov. 2005;4(10):815–8.

    CAS  Article  PubMed  Google Scholar 

  14. 14.

    Bagherzadeh K, Shirgahi Talari F, Sharifi A, Ganjali MR, Saboury AA, Amanlou M (2014) A new insight into mushroom tyrosinase inhibitors: docking, pharmacophore-based virtual screening, and molecular modeling studies. J Biomol Struct Dyn. (ahead-of-print):1-15.

  15. 15.

    Abbasi M, Ramezani F, Elyasi M, Sadeghi-Aliabadi H, Amanlou M. A study on quantitative structure–activity relationship and molecular docking of metalloproteinase inhibitors based on L-tyrosine scaffold. DARU Journal of Pharmaceutical Sciences. 2015;23(1):1–10.

    CAS  Article  Google Scholar 

  16. 16.

    Hansch C, Kurup A, Garg R, Gao H. Chem-bioinformatics and QSAR: a review of QSAR lacking positive hydrophobic terms. Chem Rev. 2001;101(3):619–72.

    CAS  Article  PubMed  Google Scholar 

  17. 17.

    Kroemer RT. Structure-based drug design: docking and scoring. Current Protein and Peptide Science. 2007;8(4):312–28.

    CAS  Article  PubMed  Google Scholar 

  18. 18.

    Aparoy P, Reddy KK, Reddanna P. Structure and ligand based drug design strategies in the development of novel 5-LOX inhibitors. Curr Med Chem. 2012;19(22):3763.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  19. 19.

    Frisch M, Trucks G, Schlegel H, Scuseria G, Robb M, Cheeseman J, et al. Gaussian 98, revision a. 7. Pittsburgh: Gaussian, Inc; 1998. p. 12.

    Google Scholar 

  20. 20.

    Morris GM, Huey R, Lindstrom W, Sanner MF, Belew RK, Goodsell DS, et al. AutoDock4 and AutoDockTools4: automated docking with selective receptor flexibility. J Comput Chem. 2009;30(16):2785–91.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  21. 21.

    Bruncko M, Tahir SK, Song X, Chen J, Ding H, Huth JR, et al. N-aryl-benzimidazolones as novel small molecule HSP90 inhibitors. Bioorg Med Chem Lett. 2010;20(24):7503–6.

    CAS  Article  PubMed  Google Scholar 

  22. 22.

    Azizian H, Bahrami H, Pasalar P, Amanlou M. Molecular modeling of helicobacter pylori arginase and the inhibitor coordination interactions. J Mol Graph Model. 2010;28(7):626–35.

    CAS  Article  PubMed  Google Scholar 

  23. 23.

    Cosconati S, Forli S, Perryman AL, Harris R, Goodsell DS, Olson AJ. Virtual screening with AutoDock: theory and practice. Expert Opin Drug Discovery. 2010;5(6):597–607.

    CAS  Article  Google Scholar 

  24. 24.

    Morris GM, Goodsell DS, Halliday RS, Huey R, Hart WE, Belew RK, et al. Automated docking using a Lamarckian genetic algorithm and an empirical binding free energy function. J Comput Chem. 1998;19(14):1639–62.

    CAS  Article  Google Scholar 

  25. 25.

    Abbasi M, Sadeghi-Aliabadi H, Hassanzadeh F, Amanlou M. Prediction of dual agents as an activator of mutant p53 and inhibitor of Hsp90 by docking, molecular dynamic simulation and virtual screening. J Mol Graph Model. 2015;61:186–95.

    CAS  Article  PubMed  Google Scholar 

  26. 26.

    Abraham MJ, Murtola T, Schulz R, Páll S, Smith JC, Hess B, et al. GROMACS: high performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX. 2015;1:19–25.

    Article  Google Scholar 

  27. 27.

    Chovancova E, Pavelka A, Benes P, Strnad O, Brezovsky J, Kozlikova B, Gora A, Sustr V, Klvana M, Medek P (2012) CAVER 3.0: a tool for the analysis of transport pathways in dynamic protein structures.

    Google Scholar 

  28. 28.

    Søndergaard CR, Olsson MH, Rostkowski M, Jensen JH. Improved treatment of ligands and coupling effects in empirical calculation and rationalization of p K a values. J Chem Theory Comput. 2011;7(7):2284–95.

    Article  PubMed  Google Scholar 

  29. 29.

    Tropsha A, Gramatica P, Gombar VK. The importance of being earnest: validation is the absolute essential for successful application and interpretation of QSPR models. QSAR & Combinatorial Science. 2003;22(1):69–77.

    CAS  Article  Google Scholar 

  30. 30.

    Consonni V and Todeschini R. Molecular discriptors. In: Puzyn T, Leszczynski J, Cronin MT editors. Recent advances in QSAR studies: methods and applications, vol 8. London: Springer Science & Business Media; 2010. p.29-93.

  31. 31.

    Fernández M, Caballero J. Modeling of activity of cyclic urea HIV-1 protease inhibitors using regularized-artificial neural networks. Bioorg Med Chem. 2006;14(1):280–94.

    Article  PubMed  Google Scholar 

  32. 32.

    Stebbins CE, Russo AA, Schneider C, Rosen N, Hartl FU, Pavletich NP. Crystal structure of an Hsp90–geldanamycin complex: targeting of a protein chaperone by an antitumor agent. Cell. 1997;89(2):239–50.

    CAS  Article  PubMed  Google Scholar 

  33. 33.

    Prodromou C, Roe SM, O'Brien R, Ladbury JE, Piper PW, Pearl LH. Identification and structural characterization of the ATP/ADP-binding site in the Hsp90 molecular chaperone. Cell. 1997;90(1):65–75.

    CAS  Article  PubMed  Google Scholar 

  34. 34.

    Obermann WM, Sondermann H, Russo AA, Pavletich NP, Hartl FU. In vivo function of Hsp90 is dependent on ATP binding and ATP hydrolysis. J Cell Biol. 1998;143(4):901–10.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

Download references


We gratefully acknowledge the financial support from Tehran University of Medical Sciences.


Research Council of Tehran University of Medical sciences, Tehran, Iran.

Availability of data and materials

Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.

Authors’ contributions

MA: Participated in the design of the study, carried out the QSAR study, performed molecular docking, molecular dynamic simulation and drafted the manuscript. HAS: Participated in the design of the study drafted the manuscript and also provide financial and administrative support. MA: Was responsible for the study design and also gave final approval of the version to be published. All authors read and approved the final manuscript.

Competing interests

The authors declare that they have no competing interests.

Consent for publication

Not applicable.

Ethics approval and consent to participate

Not applicable.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Author information



Corresponding authors

Correspondence to Hojjat Sadeghi-Aliabadi or Massoud Amanlou.

Rights and permissions

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (, which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated.

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Abbasi, M., Sadeghi-Aliabadi, H. & Amanlou, M. Prediction of new Hsp90 inhibitors based on 3,4-isoxazolediamide scaffold using QSAR study, molecular docking and molecular dynamic simulation. DARU J Pharm Sci 25, 17 (2017).

Download citation


  • Hsp90
  • Inhibitor
  • 3,4-Isoxazolediamide
  • QSAR
  • Molecular docking
  • Molecular dynamic simulation