Discovery of new heat shock protein 90 inhibitors using virtual co-crystallized pharmacophore generation

Abstract

The pharmacophoric features of the virtual cocrystallized protein of 178 Hsp90 proteins were obtained from the protein data bank and explored to generate 1260 pharmacophores evaluated using the decoy list composed of 1022 compounds. Accordingly, 51 pharmacophores were selected with high receiver operating characteristic (ROC) value for further processing. Subsequently, genetic algorithm and multiple linear regression analysis were employed to select an optimal combination of pharmacophoric models and 2D physicochemical descriptors capable of accessing a self-consistent quantitative structure-activity relationship (QSAR) of optimal predictive potential (R₆₇² = 0.819, F = 43.0, R²_LOO= 0.782, R²_PRESS against 16 external test inhibitors equal 0.735). Two orthogonal pharmacophores emerged in the QSAR equation suggesting the existence of at least two binding modes accessible to ligands within the Hsp90 binding pocket. The fifth generated pharmacophoric model from Hsp90 protein 2XJX (2XJX_2_05), and the forth generated cocrystallized pharmacophoric model from Hsp90 protein 4LWF (4LWF_2_04) with area under the curve AUC–ROC values 0.812 and 0.876, respectively were selected to be used as a searching tool sequentially of the National Cancer Institute (NCI) database. The captured hits were mapped based on successful hypotheses and the best predicted hits were selected. Twenty-four hits showed Hsp90 inhibition, 15 hits were measured with low micromolar IC₅₀ ranged from 5.0 μM to 77.1 μM

Keywords:

• Co-crystallized structure
• cancer
• docking
• Hsp90
• pharmacophore
• NCI

View correction statement:

Correction to: Discovery of new heat shock protein 90 inhibitors using virtual co-crystallized pharmacophore generation

Introduction

Hsp90 is a molecular chaperone that plays crucial role in the conformational maturation, stability, and function of protein substrates within the cell1–6. The interaction of ATP with its binding domain in Hsp90 leads to autophosphorylation of certain tyrosine residues, thus activating this kinase and provides the necessary energy for the refolding of denatured proteins1–6. Previously screened hits show Hsp90 inhibitory effect7–11. Accordingly, we initiated an exploratory effort to evaluate a generated pharmacophore from co-crystallized structure, then the mapped compounds were assayed using malachite green assay12,13. The validity of Hsp90 as anticancer target for drug discovery was established by emerging clinical trials employing potent Hsp90 inhibitor 17-allylaminogeldanamycin and the natural Hsp90 inhibitors geldanamycin and radicicol14–17. However, despite the high cellular activity and clinical progression of 17-allylaminogeldanamycin18, it has several limitations, for example, poor solubility, hepatotoxicity and extensive metabolism. These issues have led to significant efforts to identify novel rationally designed small molecular inhibitors of Hsp9019. The main focus of recent efforts toward the development of new Hsp90 inhibitors concentrate on structure-based ligand design20–22 and high-throughput screening23, with a few ligand-based examples24,25. The continued interest in designing new Hsp90 inhibitors and lack of adequate ligand-based computer-aided drug discovery efforts combined with the drawbacks of structure based design and the significant induced fit flexibility observed for Hsp90 prompted us to explore the possibility of developing a pharmacophoric model derived from virtual Hsp90 cocrystallized structures26.

Experimental section

Molecular modeling

Pharmacophore mapping were performed using CATALYST (HYPOGEN module) and CERIUS2 software suites implemented in Discovery Studio 4.5 from Accelrys Inc. (San Diego, CA, www.accelrys.com). Structure drawing was performed employing ChemDraw Ultra 7.0 (Cambridge Soft Corp., Cambridge, MA, http://www.cambridgesoft.Com).

Receptor-ligand pharmacophore generation

Structure-based pharmacophore model utilizes the interactions between receptor-ligand complexes to generate a hypothesis27. As deposit of X-ray crystal structures in PDB is growing rapidly, the structure-based methods have become increasingly important. The information about the protein structure is a good source to bring forth the structure-based pharmacophore and used as first screening before docking studies. Hundred and seventy-six crystal structure had been selected for Hsp90 to explore the pharmacophoric features of cocrystallized ligand28. The crystal structure was downloaded from the Protein Data Bank (PDB) (http://www.rcsb.org). Hydrogen atoms were added to the protein, utilizing DS 4.5 templates for protein residues. Receptor-ligand pharmacophore generation algorithm was used within DS 4.5. The used parameters were set as follows: minimum number of features was 4 while the maximum number of features was 10, the number of generated pharmacophoric models was 10 and the cocrystallized water was kept in the protein structure. The Receptor-Ligand Pharmacophore Generation protocol of Accelrys Discovery Studio v4.5 (DS), Accelrys, San Diego, CA, was applied to accomplish this task with default parameters. This protocol generates selective pharmacophore models based on receptor–ligand interactions. First, a set of features from the binding ligand is identified. The following predefined feature types are considered: hydrogen bond acceptor (HBA), hydrogen bond donor (HBD), hydrophobic (Hbic), negative ionizable (NI), positive ionizable (PI), ring aromatic (RA)29,30. Second, the pharmacophore models were ranked based on a measure of selectivity score and the top models are returned. Third, ROC analysis of the generated Hsp90 pharmacophores; pharmacophores with good ROC analysis were selected for further QSAR analysis31. Fourth, the successful QSAR equation is used to predict the activity of captured hits that have been captured by successful pharmacophores in QSAR equation32^,33. Fifth, the best predicted hits were ranked according to the predicted activity, ordered and in vitro analyzed12.

Molecular docking

The 3D coordinates of Hsp90 PDB code: 2XJX and 4LWF were retrieved from PDB and the Gasteiger–Marsili charges were assigned to the protein atoms, as implemented within DS 4.534. The protein structures were utilized in subsequent docking experiments without energy minimization. Explicit water molecules were retained according to the required docking conditions (i.e. docking in the presence of explicit water molecules). Docking settings LigandFit considers the flexibility of the ligand and considers the receptor to be rigid according to the LigandFit algorithm35. The Monte Carlo search parameters were as follows: number of trials 15 000; search step for torsions with polar hydrogens = 30.0°. The RMS threshold for ligand-to-binding-site shape matching was set to 2.0 Å, employing a maximum of 1.0 binding-site partitions. The interaction energies were assessed employing the CFF force field (v.1.02) with a nonbonded cutoff distance of 10.0 Å and distance-dependent dielectric. An energy grid extending 3.0 Å from the binding site was implemented. The interaction energy was estimated with a trilinear interpolation value using soft potential energy approximations. Rigid body ligand minimization parameters: 20 steepest descent iterations followed by 40 BFGS-minimization iterations were applied to every orientation of the docked ligand. The best 10 poses were further energy minimized within the binding site for a maximum of 200 rigid body iterations36.

Ligand pharmacophore mapping

The Fit value for any compound is dependent on summation of mapped hypothesis features that represents the number of pharmacophore features which successfully superimpose the corresponding chemical moieties within the fitted compound30,32,35 Mapping of the compound with the pharmacophore were calculated as fit value, the fit value varies according to the mapping. In order to compare the mapped features with the docked pose, the screened NCI hits were fitted against the virtual cocrystallized Hsp90 pharmacophore models 2XJX_2_05, 4LWF_2_04 using the “best fit” option within CATALYST in DS 4.529, then the mapped compounds were docked inside the ATP binding site of Hsp90 using the rigid-body docking feature. In order to compare the mapped feature with the docked pose, the screened hits were fitted against virtual cocrystallized Hsp90 pharmacophoric models using the “best fit” option within CATALYST in DS 4.5 (29) then the mapped compounds were docked in the ATP-binding site of Hsp90 using rigid docking feature35.

ROC analysis

1260 generated pharmacophores were validated by assessing their abilities to selectively capture diverse Hsp90 active compounds from a large testing list of actives and decoys. The testing list was prepared as described by Verdonk and coworkers37,38. Diverse active compounds were selected based on hydrophilic and Hbic properties, subsequently, for each active compound in the test set; around 30 decoys were randomly chosen from the ZINC database38. The decoys were selected in such a way that they share similar properties toward their corresponding active compound. Active testing compounds were defined as those possessing Hsp90 affinities ranging from 0.0006 to 6.0 μM. The test set included 33 active compounds and 989 ZINC decoys. The test set (1022 compounds) was screened by each particular pharmacophore employing the “Best flexible search” option implemented in CATALYST, while the conformational spaces of the compounds were generated employing the “Fast conformation generation option” implemented in CATALYST29. Compounds missing one or more features were discarded from the hit list. In-silico hits were scored employing their fit values.

The success of a particular virtual screening workflow was based on the area under the ROC curve (AUC)37,38. In an optimal ROC curve, an AUC value of 1 is obtained; however, random distributions cause an AUC value of 0.5. Virtual screening that performs better than a random discrimination of actives and decoys retrieve an AUC value between 0.5 and 1, whereas an AUC value lower than 0.5 represents the unfavorable case of a virtual screening method that has a higher probability to assign the best scores to decoys than to actives37,38.

QSAR modeling

QSAR modeling commenced by selecting a subset of 67 compounds from the total list of inhibitors39–41 (1–83, Figure A, Table B under Supplementary Materials) as a training set for QSAR modeling; the remaining 16 molecules (ca. 20% of the dataset) were employed as an external test subset for validating the QSAR models. The test molecules were selected as follows: all 83 inhibitors were ranked according to their IC₅₀ values, and then every fifth compound was selected for the test set starting from the high-potency end. The selected test molecules should represent similar range of biological activities to that of the training set. The selected test inhibitors are marked with asterisks in Table B under Supplementary Materials. The logarithm of measured IC₅₀ (μM) values was used in QSAR, thus correlating the data linear to the free energy change. Subsequently, we implemented genetic algorithm and multiple linear regression analyses to select optimal combination of pharmacophoric models and other physicochemical descriptors capable of self-consistent and predictive QSAR model31.

In silico screening for new Hsp90 inhibitors

The captured pharmacophore in QSAR equation (2XJX_2_05 and 4LWF_2_04) were employed as 3D search queries sequentially to screen two 3D flexible structural databases, NCI. The screening was done employing the “Best Flexible Database Search” option implemented within CATALYST. NCI hits were filtered according to Lipinski’s42 and Veber’s41 rules. Remaining hits were fitted against the three pharmacophores using the “best fit” option within CATALYST. The fit values together with the relevant molecular descriptors of each hit were substituted in the optimal QSAR Equation (1). The highest ranking molecules based on QSAR predictions were acquired and tested in vitro.

In vitro experimental studies

Reagents and reference samples

Active Hsp90 enzyme 10 μg (Promega Corporation, Madison, WI), malachite green 99% (Sigma-Aldrich, Darmstadt, Germany), Ammonium molybdate 95% (Sigma-Aldrich, Darmstadt, Germany), Sodium citrate 99% (Sigma-Aldrich, Darmstadt, Germany), ATP 1 mM solution (Sigma-Aldrich, Darmstadt, Germany), (CCT018159) Hsp90 standard inhibitor 10 mg (Tocris Bioscience, Bristol, UK), 40 NCI compounds from National Cancer Institute (Rockville, MD).

Preparation of hit compounds for in vitro assay

The tested compounds were provided as dry powders in variable quantities (5–10 mg). They were initially dissolved in DMSO to give stock solutions of 0.02 M. Subsequently, they were diluted to the required concentrations with deionized water for enzymatic assay.

Quantification of Hsp90 activity in a spectrophotometric assay

The ATPase activity of Hsp90 was quantified by colorimetric measurement of released inorganic phosphate12,13. Bioassays were performed as follows; in a 96-well clear plate, the reaction solution of total volume of 50 μL contains 100 mM Tris/HCl, pH 7.4, 6 mM MgCl₂, 20 mM KCl, 100 μM ATP, 0.1 mg/ml BSA and 50 ng/well of human Hsp90 enzyme, 5 μL of tested compounds. The plate was sealed and the reaction was incubated at 37 °C for 24 h. The reaction was stopped by the addition of 50 μL of previously prepared malachite green solution (5.2% ammonium molybdate in H2SO4, 0.0812% malachite green, 2.32% polyvinyl alcohol and water in ratios of 1:2:1:2, respectively), followed by 10 μL of 10% Sodium citrate, left for 20 min and the absorbance at 630 nm was measured using a plate reader (Bio-Tek instruments ELx 800, Winooski, VT). The calibration curve was prepared using five different concentrations of phosphate ion (10–200 μM). The final concentration of DMSO did not exceed 1.0%. Inhibition of Hsp90 was calculated as percent activity of the uninhibited ATPase control. CCT018159 was tested as positive control, while negative controls were prepared by adding the substrate after reaction termination.

Results

Exploration of Hsp90 pharmacophoric space

A total of 178 Hsp90 proteins were used in this study26. We decided to explore the pharmacophoric space of Hsp90 inhibitors through 178 Receptor-Ligand Pharmacophore Generation27. Models were restricted to explore pharmacophoric models incorporating the following features; (PosIon), (HBA), (HBD), (Hbic), and (RingArom) features, the input features were reasonably selected based on cocrystallized ligand. Furthermore, we instructed the software to explore from 4 to 10-featured pharmacophores, and the water of crystallization was kept within the protein during the pharmacophore generation27.

Eventually, 1260 pharmacophore models emerged from 178 automatic runs. The generated models were evaluated by ROC analysis lead to 51 successful pharmacophore with acceptable score properties ranged from 0.912 to 0.802 as shown in the supplementary material in Table A. These successful models were used in subsequent QSAR modeling. Interestingly, the representative models shared comparable features and acceptable statistical success criteria.

Emergence of several statistically comparable pharmacophore models suggests the ability of Hsp90 ligands to assume multiple pharmacophoric binding modes within the binding pocket. Therefore, it is quite challenging to select any particular pharmacophore hypothesis as a sole representative of the binding process.

QSAR analysis

Despite the excellent value of pharmacophoric hypotheses in probing ligand-macromolecule recognition and as 3D search queries to search for new biologically interesting scaffolds, their predictive value as 3D-QSAR models is generally hampered by steric shielding and bioactivity-enhancing or reducing auxiliary binding groups (e.g. the biological effects of electron-donating and withdrawing substitutions)7. Moreover, our pharmacophore exploration of Hsp90 inhibitors furnished several binding hypotheses of comparable success criteria, which makes it very hard to select any particular pharmacophore as sole representative of ligand binding within Hsp90. Accordingly, we were prompted to employ classical QSAR analysis to search for the best combination of pharmacophore(s) and other 2D descriptors capable of explaining bioactivity variation across the whole list of successful pharmacophores (ROC-AUC > 0.800). That is, we employed GFA-based QSAR as competition arena to select the best pharmacophore(s), i.e. among the resulting population of binding models, and supplement it(them) with two-dimensional (2D) descriptors to correct for the weaknesses of pharmacophore models (steric shielding and bioactivity-enhancing or reducing auxiliary-binding groups). We employed genetic function approximation and multiple linear regression QSAR (GFA-MLR-QSAR) analysis to search for an optimal QSAR equation(s).

Equation (1) shows the details of the optimal QSAR model. (1) where, R²₆₇ is the correlation coefficient against 67 training compounds, R²_LOO is the leave-one-out correlation coefficient and R²_PRESS is the predictive R² determined for the 16 test compounds33. Successful models (2XJX_2_05) and (4LWF_2_04) were well described in Tables 1–4, Table 1 shows the 10 generated models from the selected protein, number of features and selectivity score, Tables 2–4 showed the model properties; pharmacophoric features and corresponding weights, tolerances of the two successful 3D-model, while Table 4 represents the fit values of the tested compounds against these pharmacophores. 2D descriptors; AlogP; Log of the partition coefficient, HBA_Count is 2D structural properties that count hydrogen bond acceptors; Jurs_RPCS (Jurs relative positive charge surface area) and Jurs_TPSA (Jurs-Total-Polar-Surface-Area) are spatial descriptor that reflect the Jurs charged partial surface area discriptors; Jurs_RPCS is defined as a relative positive charge surface area in comparison to the most positive surface area while Jurs_TPSA is defined as a sum of solvent accessible surface areas of atoms with absolute value of partial charges29. The predicted activity is dependent on different factors such as lipophilicity, number of hydrogen bond acceptors and relative positive charge distribution which reflect the nature of the Hsp90-binding site.

Table 1. Pharmacophores generated from the co-crystallized Hsp90 structure (2XJX and 4LWF).

Pharmacophore	Model	Number of features	Feature set*	Selectivity score
2XJX_2_01	Pharmacophore_01	5	DHHHP	9.7772
2XJX_2_02	Pharmacophore_02	5	AHHHP	8.8637
2XJX_2_03	Pharmacophore_03	4	DHHP	8.2624
2XJX_2_04	Pharmacophore_04	4	DHHP	8.2624
2XJX_2_05	Pharmacophore_05	4	DHHP	8.2624
2XJX_2_06	Pharmacophore_06	4	HHHP	7.3489
2XJX_2_07	Pharmacophore_07	4	AHHP	7.3489
2XJX_2_08	Pharmacophore_08	4	AHHP	7.3489
2XJX_2_09	Pharmacophore_09	4	AHHP	7.3489
2XJX_2_10	Pharmacophore_10	4	DHHH	6.8689
4LWF_2_01	Pharmacophore_01	5	ADDDH	10.622
4LWF_2_02	Pharmacophore_02	5	ADDDH	10.622
4LWF_2_03	Pharmacophore_03	5	AADDH	9.7088
4LWF_2_04	Pharmacophore_04	5	AADDH	9.7088
4LWF_2_05	Pharmacophore_05	5	AADDH	9.7088
4LWF_2_06	Pharmacophore_06	5	AADDH	9.7088
4LWF_2_07	Pharmacophore_07	4	ADDD	9.1076
4LWF_2_08	Pharmacophore_08	4	DDDH	9.1076
4LWF_2_09	Pharmacophore_09	4	ADDD	9.1076
4LWF_2_10	Pharmacophore_10	5	AAADH	8.7953

Bolded pharmacophores are the QSAR selected pharmacophores used as a searching queries.

*A: hydrogen bond acceptor; D: hydrogen bond donor; H: hydrophobe; P: positive ionizable; R: ring aromatic.

Table 2. Pharmacophoric features and corresponding weights, tolerances and 3D coordinates of generated pharmacophore from cocrystallized Hsp90 (2XJX_2_05).

		Chemical features
Model	Definition		Hbic1	Hbic2	HBD		PosIon
2XJX_2_05	Tolerances		1.60	1.60	1.60	2.20	1.60
	Coordinates	X	13.68	15.88	11.32	9.702	14.297
		Y	0.22	−5.6	−2.163	− 4.06	−4.266
		Z	9.48	3.86	11.12	12.786	0.628
Feature	X	Y	Z	Radius
ExcludedVolume_1.6	17.99	−9.955	1.538	1.27
ExcludedVolume_1.7	14.379	−9.186	6.516	1.368
ExcludedVolume_1.8	11.584	4.072	10.408	1.6
ExcludedVolume_1.9	11.629	−0.727	15.258	1.154
ExcludedVolume_1.10	17.657	1.522	4.208	1.27
ExcludedVolume_1.11	19.143	−9.665	7.247	1.454
ExcludedVolume_1.12	11.109	−7.504	3.428	1.53
ExcludedVolume_1.13	19.786	−6.747	9.056	1.27
ExcludedVolume_1.14	16.369	−5.262	13.14	1.368
ExcludedVolume_1.15	19.648	−2.737	9.129	1.454
ExcludedVolume_1.16	15.887	1.055	14.295	1.154
ExcludedVolume_1.17	16.132	3.478	9.757	1.154
ExcludedVolume_1.18	13.118	−8.426	12.278	1.27
ExcludedVolume_1.19	10.355	−3.364	7.132	1.53
ExcludedVolume_1.20	10.349	−7.583	9.559	1.454
ExcludedVolume_1.21	7.465	−2.365	11.724	1.454
ExcludedVolume_1.22	4.726	−8.104	−27.842	1.154
ExcludedVolume_1.23	−5.074	−14.009	−26.614	1.53
ExcludedVolume_1.24	−0.172	−11.579	−30.072	1.454

Table 3. Pharmacophoric features and corresponding weights, tolerances and 3D coordinates of generated pharmacophore from cocrystallized Hsp90 (4LWF_2_04).

		Chemical features
Model	Definition		HBD1		HBD2		Hbic
4LWF_2_04	Tolerances		1.60	2.20	1.60	2.20	1.60
	Coordinates	X	−4.016	−6.446	−0.634	− 0.653	1.76
		Y	−10.343	−11.582	−7.315	− 6.949	−12.38
		Z	−24.81	−23.561	−28.988	− 31.966	−23.5
			HBA1		HBA2
	Tolerances		1.60	2.20	1.60	2.20
	Coordinates	X	−1.273	0.808	−2.269	− 4.493
		Y	−12.44	−10.723	−9.029	− 10.62
		Z	−21.476	−20.164	−28.366	− 27.132
Feature	X	Y	Z	Radius
ExcludedVolume_1.6	17.99	−9.955	1.538	1.27
ExcludedVolume_1.7	14.379	−9.186	6.516	1.368
ExcludedVolume_1.8	11.584	4.072	10.408	1.6
ExcludedVolume_1.9	11.629	−0.727	15.258	1.154
ExcludedVolume_1.10	17.657	1.522	4.208	1.27
ExcludedVolume_1.11	19.143	−9.665	7.247	1.454
ExcludedVolume_1.12	11.109	−7.504	3.428	1.53
ExcludedVolume_1.13	19.786	−6.747	9.056	1.27
ExcludedVolume_1.14	16.369	−5.262	13.14	1.368
ExcludedVolume_1.15	19.648	−2.737	9.129	1.454
ExcludedVolume_1.16	15.887	1.055	14.295	1.154
ExcludedVolume_1.17	16.132	3.478	9.757	1.154
ExcludedVolume_1.18	13.118	−8.426	12.278	1.27
ExcludedVolume_1.19	10.355	−3.364	7.132	1.53
ExcludedVolume_1.20	10.349	−7.583	9.559	1.454
ExcludedVolume_1.21	7.465	−2.365	11.724	1.454
ExcludedVolume_1.10	0.475	−17.398	−25.583	1.154
ExcludedVolume_1.11	−1.187	−3.518	−27.257	1.154
ExcludedVolume_1.12	4.565	−12.764	−26.382	1.53

Table 4: NCI compounds with Hsp90 inhibitory values.

No	NCI*	2XJX_2_05†	4LWF_2_04‡	%Inhibition¶	IC50 μM	R^2
84	365351	2.772	2.551	94.7	5.0	0.997
85	687759	0.614	1.83	100	8.2	0.984
86	120626	3.241	1.441	100	9.0	0.957
87	180024	1.214	3.442	100	10.1	0.933
88	179203	2.723	2.016	100	11.3	0.994
89	123468	2.666	1.198	100	14.0	0.90
90	276751	1.847	1.668	79.0	15.8	0.991
91	736219	1.506	3.618	100	16.0	0.90
92	122060	2.559	1.936	100	18.9	0.981
93	111066	3.097	1.699	70.5	19.2	0.975
94	120924	1.072	1.346	87.5	22.0	0.999
95	112531	2.991	1.311	77.3	29.5	0.964
96	125844	2.868	1.682	73.8	40.2	0.90
97	736218	1.414	2.258	74.5	48.3	0.977
98	112534	3.187	1.181	53.4	77.1	0.993
99	134420	2.275	1.132	86.3	--	--
100	637992	2.921	1.193	57.9	--	--
101	268696	2.514	3.194	52.6	--	--
102	268708	0.634	2.657	51.3	--	--
103	124232	2.416	1.773	50	--	--
104	143491	3.546	2.862	48.7	--	--
105	112530	3.179	1.643	38.6	--	--
106	125845	1.292	1.599	21.3	--	--
107	652604	1.544	2.732	17.1	--	--
108	CCT018159^d	----	----	98.2	7.2	0.97

*NCI screened and bioassay hits.

† Hsp90 pharmacophoric model generated from cocrystallized structure.

‡ Percentage of inhibition at 100 μM concentration of inhibitor.

¶ Standard Hsp90 inhibitor with reported IC₅₀=5.7 μM42.

ROC analysis

Hundred and seventy-eight cocrystallized Hsp90 protein were used to generate 1260 pharmacophores. The generated pharmacophores were analyzed and sorted by ROC analysis using decoy list (1022 compounds) yield 51 pharmacophores with AUC > 0.800 and evaluated as good pharmacophore as shown in supplementary materials in Table A. QSAR analysis of the best 51 generated pharmacophores lead to a selection of two pharmacophores (2XJX_2_05, and 4LWF_2_04) (Tables 2, 3 and Figures 1A, 2A). Both pharmacophores were used as a searching tool sequentially for NCI database (257 000 compounds), out of which 88 compounds were captured and ranked according to the fit values.

Figure 1. (A) 2XJX_2_05, the QSAR successful pharmacophore generated by DS studio 4.5, (B) cocrystallized XJX1232 in Hsp90 (2XJX, resolution 1.66 Å), (C) mapping of pharmacophore with cocrystallized Ligand XJX1232, (D) chemical structure of XJX1232.

Figure 2. (A) 4LWF_2_04, the QSAR successful pharmacophore generated by DS studio 4.5, (B) cocrystallized FJ3301 in Hsp90 (4LWF, resolution 1.75 Å), (C) mapping of pharmacophore with cocrystallized Ligand FJ3301, (D) chemical structure of FJ3301.

Virtual screening and in vitro validation

The highest ranked 40 compounds were ordered and tested in vitro for Hsp90 inhibition using malachite green assay12,13. Twenty-four tested compounds (84–107) (Figures 3) showed Hsp90 percentage of inhibition at 100 μM with the fit values as illustrated in Table 4.

Figure 3. Chemical structure of tested NCI compounds.

Figure 4. Two dimensional analysis of the cocrystallized ligand in the ATPase binding site of (A) Hsp90 (2XJX , resolution 1.66 Å) (B) Hsp90 (4LWF , resolution 1.75 Å), using DS 3.5 visualizer.

Figure 5. (A, D, G) Mapping compounds (84, 86, 88), respectively with 2XJX_2_05; cocrystallized pharmacophore generated by DS studio 4.5, (B, E, H) docked compounds (84, 86, 88), respectively in ATPase binding pocket of Hsp90 (2XJX, resolution 1.66 Å), (C, F, I) compounds (84, 86, 88) mapped and docked.

Discussion

Figure 1 shows the interaction of cocrystallized Ligand XJX1232 with the amino acids in the binding pocket; ASP93, LYS58, MET98, ASN51, LYS112, PHE138, LEU107, THR184, VAL150, ALA55, ILE96 and also with the bridging water HOH2043 and HOH2042. These interactions are expressed by (HBD), (PosIon) and (Hbic) features. It is clear that forbidden areas for interactions are filled with exclusion spheres which represent the steric clash and unfavorable interactions. Furthermore, Figure 4(A) shows the possible interaction of cocrystallized structure (XJX1232) inside the binding pocket of Hsp90 (pdb: 2XJX) after 2D analysis using DS3.5 visualizer. It shows the cocrystallized compound (XJX1232) inside the ATPase binding site of Hsp90, the corresponding amino acids that are involved in the interactions significantly are; ASP93, ASN51, THR184, ASP54, LEU48 (Hydrogen bond interactions) in addition to MET98, LEU107, VAL150, LYS58, ALA55, VAL186 and ILE96 (Van der Waal interactions–Hbic interactions). Those amino acids form the binding pocket for the corresponding cocrystallized ligand (XJX1232). By comparison the mapping features of 2XJX_2_05 in Figure 1(C); that could be explained as follows; isopropyl group mapped with Hbic feature corresponding to Van der Waal interaction with Hbic part of VAL150, LEU107, LEU48, VAL186, resorcinolic OH group mapped with HBD corresponding to interaction of OH groups with carbonyl group of LEU48. Furthermore, aromatic group mapped with Hbic feature corresponding to interaction of fused phenyl group in isoindole with Hbic pocket of the following amino acids; ALA55, ILE96 and LYS58, piperidine ring is mapped with PI feature corresponding to interaction of positively ionizable nitrogen atom with the negatively ionized carboxylate of ASP54 as shown in Figure 1(B). Mapping of cocrystallized ligand XJX1232 with model 2XJX_2_05 as shown in Figure 1(C) correlated with the cocrystallized pose within the binding pocket of Hsp90 (PDB: 2XJX, resolution 1.66 Å). Figure 5 showed docked poses of compounds 84, 86 and 88 inside the binding pocket of Hsp90 compared with the mapped features in 2XJX_2_05 as follows: sulfur atom in the thiazole heterocycle in compound 84, dimethyl group in compounds 86 and phenyl group in compound 88 mapped with Hbic feature corresponding to Van der Waal interaction with Hbic part of ALA55, LYS58, ILE96 and MET98, HBD corresponding to interaction of NH2 groups of compound 84, amidic NH group of compound 86 and NH₂ of guanidino group of compound 88 with corresponding carbonyl groups of LUE48, PI nitrogen of triazine group in compound 86 and nitrogen atom of guanidino group in compound 84 and 88 interact through ionic bond with negatively charged carboxylate group of ASP54 corresponding to mapping with (PosIon) feature in model 2XJX_2_05 as shown in Figure 5.

Similarly, Figure 2 shows the interaction of cocrystallized ligand FJ3301 with the amino acids in the binding pocket; ASP93, LYS58, MET98, ASN51, PHE138, LEU107, and also with the bridging water HOH403, HOH404 and HOH474. Furthermore, Figure 4(B) shows the possible interaction of cocrystallized structure (FJ3301) inside the binding pocket of Hsp90 (pdb: 4LWF) after 2D analysis using DS3.5 visualizer. It shows the cocrystallized compound (FJ3301) inside the ATPase-binding site of Hsp90, besides the corresponding amino acids that are involved in the interactions significantly such as; ASP93, ASN51, THR184, GLY97, LEU48, SER52 (Hydrogen bond interactions) in addition to MET98, LEU107, VAL150, PHE138, ALA55, VAL186 and ILE96 (Van der Waal interactions–Hbic interactions). Those amino acids form the binding pocket for the corresponding cocrystallized ligand (FJ3301). By comparison, the mapping features of 4LWF_2_04 in Figure 2(B, C); that could be explained as follows: isopropyl group mapped with Hbic feature corresponding to Van der Waal interaction with Hbic part of LEU107, PHE138, one of resorcinolic OH group mapped with HBD corresponding to interaction of OH groups with carboxylic acid group of ASP93 and bridging water HOH403, while the another resorcinolic OH group mapped with HBA corresponding to interaction of OH groups with NH group of ASN51 and bridging water HOH404. Furthermore, primary amine group on isoxazole aromatic group mapped with HBD corresponding to hydrogen bonding with carbonyl group of GLY97 and bridging water HOH474, isoxazole group mapped with HBA corresponding to interaction of heterocyclic oxygen atom with OH group of THR184 and bridging water HOH403 as shown in Figure 2(B). Mapping of cocrystallized ligand FJ3301 within model 4LWF_2_04 as shown in Figure 2(C) correlated with the cocrystallized pose within the binding pocket of Hsp90 (PDB: 4LWF, resolution 1.75 Å). Figure 6 showed docked poses of compounds 84, 86 and 88 inside the binding pocket of Hsp90 compared with the mapped features in 4LWF_2_04 as follows: for compound 84; diamino group of guanidino group mapped with HBD and HBA corresponding to interaction of first NH₂ with carboxylate group of ASP102, while the second NH₂ group interact with carbonyl group of GLY97 furthermore, amino group (NH₂) substituted on thiadiazole heterocyclic ring mapped with HBD corresponding to interaction of NH₂ group with carboxylate group of ASP93, heterocyclic nitrogen of thiadiazole group mapped with HBA corresponding to hydrogen bonding with NH group of ASN51, sulfur atom mapped with Hbic feature corresponding to interaction with Hbic part of LEU107 and PHE138, as shown in Figure 6(A,B,C), similar interactions with compound 86 as shown in Figure 6(D,E,F) as follows:

Phenyl group in compound 86 mapped with Hbic feature corresponding to Van der Waal interaction with Hbic part of LEU107 and PHE138, two HBD corresponding to interaction of two NH₂ groups of compound 86, the first NH₂ group mapped with corresponding carbonyl group of GLY97 and carboxylate group of ASP102, the second NH₂ group interact through bridging water HOH403 and carboxylate group of ASP93, heterocyclic nitrogen in triazine ring in compound 86 interact through hydrogen bonding with OH group of hydrophilic amino acid THR184 and bridging water HOH403 corresponding to mapping with HBA feature in model 4LWF_2_04 as shown in Figure 6(G,H,I), similar interactions with compound 88 as shown in Figure 6(G,H,I) as follows:

Diamino group of guanidino group mapped with HBD and HBA corresponding to interaction of first NH₂ with carboxylate group of ASP102, while the second NH₂ group interact with carbonyl group of GLY97 furthermore, amino group (NH₂) of second guanidino group mapped with HBD corresponding to interaction of NH₂ group with carboxylate group of ASP93, carbonyl group of ester functional group mapped with HBA corresponding to hydrogen bonding with NH group of ASN51, phenyl group mapped with Hbic feature corresponding to interaction with Hbic part of LEU107, MET98 and PHE138, as shown in Figure 6(G,H,I).

Figure 6. (A, D, G) Mapping compounds (84, 86, 88), respectively with 4LWF_2_04; cocrystallized pharmacophore generated by DS studio 4.5, (B, E, H) docked compounds (84, 86, 88), respectively in ATPase binding pocket of Hsp90 (4LWF, resolution 1.75 Å), (C, F, I) compounds (84, 86, 88) mapped and docked.

It is clear that several types of interactions with corresponding amino acids in the binding pocket, which support the suggestion of more than one of successful binding pose of the inhibitor inside the ATPase-binding pocket of Hsp90. Cocrystallized pharmacophore generation interpret the interaction with the corresponding amino acids that can be used in future to evaluate the best structure-based models.

Conclusion

A new idea for drug design, a hybrid method between both ligand-based and structure-based models was used to find out novel and potent Hsp90 inhibitors. The generated pharmacophore models from the cocrystallized protein structure were ranked according to the ROC–AUC values; the highest ranked pharmacophore models were further underwent QSAR analysis, as different pharmacophore models generated from different cocrystallized protein structures, it may represent different binding modes, the generated pharmacophores were evaluated. After the successful validation of the developed pharmacophore models, a smart virtual screening strategy was conducted by employing the highest ranked pharmacophore model to retrieve hits with novel chemical scaffolds. The final hits were tested using the malachite green colorimetric assay that lead to the identification of 24 inhibitors, out of which 15 hits with low micromolar IC₅₀ ranged from 5.0 μM to 77.1 μM.

Declaration of interest

It is to declare that the authors have no conflict of interest.

Supplementary materials available online

Acknowledgements

The authors thank the National Cancer Institute for supporting us with NCI samples. We also thank the deanship of Scientific Research at the Zarqa University for their generous funds.