Structural ensemble-based docking simulation and biophysical studies discovered new inhibitors of Hsp90 N-terminal domain

Abstract

Heat shock protein 90 (Hsp90) is one of the most abundant cellular proteins and plays a substantial role in the folding of client proteins. The inhibition of Hsp90 has been regarded as an attractive therapeutic strategy for treating cancer because many oncogenic kinases are Hsp90 client proteins. In this study, we report new inhibitors that directly bind to N-terminal ATP-binding pocket of Hsp90. Optimized structure-based virtual screening predicted candidate molecules, which was followed by confirmation using biophysical and cell-based assays. Among the reported crystal structures, we chose the two structures that show the most favourable early enrichments of true-positives in the receiver operating characteristic curve. Four molecules showed significant changes in the signals of 2D [¹H, ¹⁵N] correlation NMR spectroscopy. Differential scanning calorimetry analysis supported the results indicating direct binding. Quantified dissociation constant values of the molecules, determined by a series of 2D NMR experiments, lie in the range of 0.1-33 μM. Growth inhibition assay with breast and lung cancer cells confirmed the cellular activities of the molecules. Cheminformatics revealed that the molecules share limited chemical similarities with known inhibitors. Molecular dynamics simulations detailed the putative binding modes of the inhibitors.

Figure 1

Structural ensemble, distributions of pose reproductions by algorithms, and structure-dependent profiles of ROC curves. (A) The ensemble of 108 crystal structures is represented with the substrate-binding pocket coloured in blue and the position of Leu-107 in red. (B) Algorithm-dependent success rates in reproducing the poses of the inhibitors found in X-ray structures are represented as histograms. The 90 crystal inhibitors were docked into the structures of Hsp90N with 6 algorithms from 5 software programs. A docked pose is judged as a success when it overlapped with that of the crystal structure within 2 Å in the heteroatoms. (C) Receiver operating characteristic (ROC) curves from two structures for high-throughput virtual screening in this study are drawn in red (2BYI-A) and magenta (2YI5-A). The blue line corresponds to that of 1UYG-A, which was used as the DUD-E benchmark. Random enrichment is shown in black for comparison. In the ROC curves, the x-axis is logarithmically scaled to emphasize the earlier enrichments of true-positives. The values of LogAUC for 2BYI-A, 2YI5-A, and 1UYG-A are 33.5, 31.9, and 12.9, respectively.

Figure 2

Quantification of inhibitor binding using 2D NMR and differential scanning fluorimetry. (A) Example of 2D [¹H, ¹⁵N] HSQC in the titrations with 1. A series of ligands with different concentrations in 1:0 to 1:2 protein:ligand ratios were added. NMR peaks of 1:0 and 1:2 ratios are coloured in green and magenta, respectively. The residues exemplified in calculating K_d are circled. (B) The simulated NMR peaks by line-shape analyses for calculating K_d of 1 are drawn for comparison with the raw data. (C) Differential scanning fluorimetry in the apo and holo states are represented. Differentials of relative fluorescence units (RFU) for temperature are plotted in blue, green, cyan, red, magenta colours for 1–4 and 17-DMAG, respectively. Black lines correspond to the profiles of apo states. Calculated differences of mid-point melting temperatures from that of apo state (44.4 °C), ΔT_ms, in 1–4 and 17-DMAG are 7.3, 8.9, 3.5, 6.5, and 11.3 degrees, respectively.

Figure 3

Mapping of perturbed residues by 2D NMR onto 3D structure. The changes of 2D [¹H, ¹⁵N] HSQC in the titrations are quantified as ΔCS using $\sqrt{{0.5\times (\mathrm{\Delta }\mathrm{H}}^2+{(0.2\times \mathrm{\Delta }\mathrm{N})}^2)}$. Here ΔH and ΔN mean the chemical shift difference between apo and holo forms in ¹H and ¹⁵N dimensions. Once calculating the mean and standard deviation (SD) values of ΔCS in the cases of ΔCS > 0, the residues of Hsp90N are classified into four criteria: (i) ΔCS ≤ mean + SD, (ii) mean + SD < ΔCS ≤ mean + 2 SD, (iii) ΔCS > mean + 2 SD, and (iv) disappeared. The residues coloured in yellow, red, and orange correspond to the cases of (ii), (iii), and (iv), respectively.

Figure 4

Cell growth inhibition by 1–4. Cell growth inhibitions were observed in (A) MCF7, (B) A549, and (C) MCF10A cells. Three experiments were repeated in a condition to generate standard deviations that are represented by the error bars. 17-DMAG was used as a positive control. (D) The amounts of Hsp70 mRNA were quantified using real-time PCR in MCF7 cells. The filled and open bars indicate the measurements in 3 and 6 h, respectively. The concentrations of the added 1–4 were 10 μM, whereas that of 17-DMAG was 100 nM.

Figure 5

Cheminformatics for 1–4. Similarity ensemble approach was employed to quantify the chemical similarities of 1–4 and known Hsp90N inhibitors. Tanimoto coefficients (Tcs) between a molecule and 617 BindingDB-deposited inhibitors were summed to form an individual ∑Tc. (A) The ∑Tc of the 617 known inhibitors represented using a histogram as reference. The mean value of the ∑Tc distribution, 8.28, is drawn as a blue line. The ∑Tc for 1–4 was also calculated and is represented as red lines. The ∑Tc for 1–4 were 1.86, 1.35, 2.19, and 0.96, respectively. Only Tcs greater than the threshold, 0.3, were considered. (B) The chemicals most closely related to 1–4 by Tc from the BindingDB-deposited Hsp90N inhibitors are shown. The corresponding hit compound, Tc, and ZINC IDs are given in parentheses.

Figure 6

Simulated poses of 1–4 in complex with Hsp90N. (A) 1–4 were docked into 2BYI-A. The hydrophilic interactions are shown with dashed lines. The underlined residues are involved in hydrophobic interactions, whereas the others are involved in hydrophilic interactions. The residues with “+” indicate those that interact with backbone atoms. The structures of Hsp90N are the same direction as that in Fig. 1. (B) Ligands similar to 1–4 in the Hsp90N complex structures, 2YJW and 2YI0, are overlaid based on their protein coordinates. The white and purple sticks correspond to 2YJW and 2YI0, respectively. The Tc value between 2YJW and 2 is 0.36, while that between 2YI0 and 3 is 0.34.

Figure 7

Molecular dynamics simulations of complex structures with 1–4. The time-dependent distances between the heteroatoms forming intermolecular hydrogen bonds are drawn from 100 ns molecular dynamics (MD) simulations. Each MD simulation of a complex with 1–4 was repeated three times with different random seeds, coloured in blue, green, and red. Asp-93′s OD2 (A) and Gly-97′s O (C) were used as hydrogen bond acceptors, whereas Thr-184′s OG1 (B) was a donor.

Figure 8

Classification of Hsp90N structures based on Leu-107. (A) Crystal structures of Hsp90N were classified based on the position of Leu-107. All structures are rotated from that shown in Fig. 1 by 60 degrees along the x-axis for clarity. Two groups are labelled as open and closed according to the Leu-107 position. (B) The distribution of LogAUCs in the structures of the two groups. Those from the open forms are drawn in white, whereas the closed forms are represented in grey.