Development of a First-in-Class Small-Molecule Inhibitor of the C-Terminal Hsp90 Dimerization

Heat shock proteins 90 (Hsp90) are promising therapeutic targets due to their involvement in stabilizing several aberrantly expressed oncoproteins. In cancerous cells, Hsp90 expression is elevated, thereby exerting antiapoptotic effects, which is essential for the malignant transformation and tumor progression. Most of the Hsp90 inhibitors (Hsp90i) under investigation target the ATP binding site in the N-terminal domain of Hsp90. However, adverse effects, including induction of the prosurvival resistance mechanism (heat shock response or HSR) and associated dose-limiting toxicity, have so far precluded their clinical approval. In contrast, modulators that interfere with the C-terminal domain (CTD) of Hsp90 do not inflict HSR. Since the CTD dimerization of Hsp90 is essential for its chaperone activity, interfering with the dimerization process by small-molecule protein–protein interaction inhibitors is a promising strategy for anticancer drug research. We have developed a first-in-class small-molecule inhibitor (5b) targeting the Hsp90 CTD dimerization interface, based on a tripyrimidonamide scaffold through structure-based molecular design, chemical synthesis, binding mode model prediction, assessment of the biochemical affinity, and efficacy against therapy-resistant leukemia cells. 5b reduces xenotransplantation of leukemia cells in zebrafish models and induces apoptosis in BCR-ABL1+ (T315I) tyrosine kinase inhibitor-resistant leukemia cells, without inducing HSR.


Introduction
The heat shock protein (Hsp90) is an abundant, cytosolic molecular chaperone that modulates the folding, stabilization, and maturation of over 400 client proteins in eukaryotes. 1 It is involved in essential processes such as signal transduction, cell cycle progression, and transcription regulation. 1 In cancer cells, Hsp90 is overexpressed, involved in uncontrolled proliferation and anti-apoptotic effects, and, that way, essential for the malignant transformation and progression of several cancer types. 2 Thus, cancer cells are more dependent on Hsp90 activity than normal cells. 3,4 In various cancers, multiple signal transduction-promoting oncoproteins are clients of Hsp90. 5 Hence, inhibiting the activity of Hsp90 is a promising strategy for the development of anticancer therapy. Several Hsp90 inhibitors (Hsp90i) have been developed recently, however adverse effects including dose-limiting ocular and cardiac toxicity and poor patient stratification have precluded their clinical approval. 3 Most of the 19 Hsp90 inhibitors (Hsp90i) studied in clinical trials so far (except RTA 901) target the N-terminal ATP binding site (NTD) and are termed N-terminal inhibitors. 3,6-15 A common pitfall of Hsp90 NTD-targeting inhibitors is the induction of a pro-survival heat shock response (HSR). 3,6 The HSR is a stress response mechanism mediated by heat-shock factor 1 (HSF-1), which leads to the expression of other heat shock proteins (HSPs) including Hsp27, Hsp40 and Hsp70 as a rescue mechanism upon Hsp90 inhibition and eventually weakens cytotoxic effects of Hsp90i. 3,6,[11][12][13][14] In addition, Hsp90 NTD-targeting inhibitors potentially inflict cytotoxicity through mechanisms that involve targets other than Hsp90, (off-target effects). 4,16 The off-target effect hypothesis is also supported by the drastic differences between cytotoxicity concentrations of Hsp90 NTD-targeting inhibitors vs. their binding affinity to Hsp90. 16 Furthermore, there are two major cytosolic isoforms of Hsp90 (Hsp90α and Hsp90β) expressed in humans. Hsp90α is an inducible isoform, overexpressed in several cancer types, whereas the Hsp90β isoform is expressed constitutively. Thus, targeting Hsp90 with isoform-specific inhibitors can afford a therapeutic window. 17,18 Most of the previously studied Hsp90i exhibit pan-inhibitory activity (i.e., targeting all isoforms). However, the Hsp90α and Hsp90β isoforms share a high degree of similarity (SI Figure S21), making it challenging to develop isoform-selective inhibitors. 18 Hsp90 is a flexible homodimer, and each monomer consists of three major functional domains: Nterminal domain (NTD), middle domain, and C-terminal domain (CTD). Hsp90's activity depends on the binding and hydrolysis of ATP at the NTD and on its dimerization via the CTD. 2 The middle domain (MD) that connects the NTD and the CTD mediates the binding of clients and cochaperons. The CTD is connected to the MEEVD motif, which interacts with the subset of tetratricopeptide repeat (TPR) domain-containing cochaperones. 2 To our knowledge, inhibiting Hsp90 dimer formation by targeting the CTD dimerization interface constitutes a so far unexplored mode of action (MOA) of small-molecule Hsp90i. In contrast to Hsp90i targeting the N-terminal ATP binding site, C-terminal inhibitors do not generally induce HSR. 6,16,[19][20][21] The most important classes of C-terminal inhibitors are: (1) inhibitors binding to the C-terminal ATP-binding site (e.g., novobiocin and analogs), (2) modulators of the Hsp90-CDC37 interaction (e.g., celastrol, induces HSR 22 ), (3) modulators of the Hsp90-p23 interaction (e.g., gedunin), (4) modulators of the Hsp90-HOP interaction (e.g., LB76), and (5) aminoxyrone (AX), the first non-peptidic inhibitor of the C-terminal dimerization of Hsp90. 23 Following a strategy recently introduced by us to identify PPI inhibitors, 24,25 we initially identified hot spot residues in the CTD dimerization interface that accounted for most of the binding affinity 26 and identified the first peptidic inhibitors shown to bind to the CTD of Hsp90. 27 Furthermore, we developed AX, the first peptidomimetic Hsp90 CTD dimerization inhibitor, 28 which is a promising lead candidate effective against BCR-ABL1+ TKI-resistant leukemic cells. 28 Based on these experiences, here, we report the rational design, chemical synthesis, binding mode model, biochemical affinity, and biological in vitro evaluation of the first-in-class small molecule inhibitor (5b) of Hsp90 CTD dimerization based on a tripyrimidonamide scaffold.

Design of tripyrimidonamides as CTD Hsp90 inhibitors
Based on computational predictions and subsequent experimental validation, we identified the spatially clustered hot spot residues I688, Y689, I692, and L696 in the Hsp90 CTD interface, which are located on α-helix H5, form a functional epitope, and account for most of the protein dimerization energy. 26 Furthermore, conformational analysis by 2D NMR and MD simulations revealed for the recently introduced tripyrimidonamide scaffold that it can act as a potential α-helix mimetic, mimicking side chains at positions i, i+4 (dimeric compound) or i, i+4, i+8 (trimeric compound). 29,30 This side chain pattern is concordant with the succession of the hot spot residues in the Hsp90 CTD interface.
Together, this provided the incentive for us to design and synthesize the tripyrimidonamide 5a, which mimics the hot spots I688, I692, and L696. In compound 5a, the side chain of V was used instead of I to avoid diasteromeres. 6, which lacks the isopropyl side chain, was also designed to probe the influence of the absence of the third side chain in a tripyrimidonamide. Next, we aimed to design compounds that can also form polar interactions, as these should confer specificity of binding. 26 In addition, the binding to a well-defined cleft or groove in a PPI region has been described to yield a particularly effective PPI inhibitor. 26 The Y689 side chain of 7a should be accommodated in an indentation in the binding epitope of helix H4' (Figure 1); 26 we also designed the homolog 7b with a prolonged (4-hydroxy-phenyl)-ethyl side chain. Although both compounds mimic the three hot spots Y689, I692, L696, with the longer side chain in 7b, we intended to accommodate for the apparent mismatch between the preferred side chain orientations in tripyrimidonamides and the side chain pattern of the hot spots (i, i+3, i+4). The side chain patterns of 5b and 7a are almost identical to that of the α-aminoxy-peptide AX, which was shown to bind to the CTD. 28 Further analysis of the physicochemical properties of the CTD dimerization interface revealed a particular hydrophobic patch there (Figure 1). Interestingly, the 4-methoxy-benzyl side chain of 5b should act as a (weak) hydrogen bond acceptor for S673' and T669' on helix H4', but at the same time decrease the side chain's hydrophilicity for a more favorable burial in the overall hydrophobic interface.
To probe this with a larger substituent, we also designed the benzyloxy derivatives 5c and 5d, respectively, which are also precursors of 7a and 7b.

Synthesis of tripyrimidonamides
The monomeric building blocks 1 and 2a-e were prepared according to our previously published protocol. 30 Subsequently, the designed tripyrimidonamides 5a-d were synthesized using a modular approach. Briefly, a COMU-mediated amide coupling of the lithium carboxylate 2e with 5-aminopyrimidone 1 afforded the benzoyl-protected dimer 3 in 75% yield. Deprotection of the benzoylgroup by treatment of 3 with sodium hydroxide in methanol at 80 °C afforded the unprotected dimer 4 (77% yield). Additional coupling reactions of 4 with the respective lithium salts 2a-d in the presence of COMU furnished the tripyrimidonamides 5a-d in 39-76% yield. Compound 6 with an N-unsubstituted N-terminal pyrimidone ring was synthesized by treating the corresponding 4-methoxybenzyl-substituted derivative 5b with BBr 3 in DCM (Scheme 1). Finally, the trispyrimidonamides 7a and 7b with free phenolic groups were prepared by catalytic hydrogenation of their respective O-benzyl-protected precursors 5c and 5d (40 and 87% yield).  31 shown in surface and cartoon representation. For one of the Hsp90 monomers, the Nterminal domain (NTD) is colored in red, the middle domain in beige, and the C-terminal domain (CTD) in blue. Above and below the protein structure, the structures of Hsp90i and their potential binding sites (see ref. [32][33][34][35][36] , color-coded according to the domains) are shown. b) Dimeric CTD of human Hsp90β with the two monomers in blue and white. Helices H4, H4', H5, and H5' of the CTDs form the dimerization interface. c) Residues forming the CTDs dimerization interface in human Hsp90α are primarily located on helices H4, H4', H5, and H5'. 26 d) Trispyrimidones can adopt conformations resembling the side chain orientation of an α-helix in i, i+4 and i+7 position. 30 e) Synthesis of tripyrimidoneamides: a) COMU, DMF, r. t., 18 h; b) NaOH, MeOH, 80 °C, 6 h; a) 2 ad, COMU, DMF, r. t., 18 h; c) 6 via 5b, BBr 3 , DCM, -78 °C, 1 h, r. t., 1 h; d) via 7a via 5c and 7b via 5d, H 2 , Pd(C), MeOH, DCM, r. t., 1 h.

Selection of 5b as a lead candidate
To evaluate the inhibition of Hsp90 dimerization, E. coli BL21 (DE3) pETSH-3 cells were used to display Hsp90α on their surface (Figure 2a). 27 Passenger-driven dimer formation of Hsp90α is facilitated through the motility of the β-barrel domain within the outer membrane of E. coli, as reported for other proteins. 37 To demonstrate the functionality of dimerized Hsp90 on the surface of E. coli, the transcription factor p53, a natural client protein of Hsp90, was labeled with fluorescein isothiocyanate (FITC) and added to cells displaying Hsp90 on their surface. Subsequent flow cytometer analysis revealed a high green fluorescence for cells displaying Hsp90, indicating dimerized and functional Hsp90 (Figure 2b). Compounds 5a, 6, and 7a showed only weak inhibition of 3.27 %, 14.65 %, and 24.35 %, respectively. In contrast, 5b, 5c, 5d, and 7b showed moderate inhibition of 39.92 %, 41.83 %, 55.23 %, and 31.33 %, respectively (Figure 2b).
Later, the binding affinity of the compounds was determined with microscale thermophoresis (MST) measurements, using NT-647-labeled recombinant CTD of Hsp90α protein. 28 A nonlinear regression curve was fitted with the K D formula, and, as expected, substances showing weak inhibition have high dissociation constants (6: 249 µM, 7a: 286 µM; Figure 2c, e). The lowest K D value was observed for 5b with 3.42 µM (Figure 2c, e). Next, the in vitro cytotoxicity assessment of compounds 5a-d, 6, 7a, and 7b revealed 5b as a promising candidate (with low IC 50 : 1.8 ± 0.3 µM) in a BCR-ABL1+ tested leukemia cell line K562 (Figure 2d, 2e).
Based on the inhibition of Hsp90α CTD dimerization, low apparent K D value for the Hsp90α CTD, and potent anti-leukemic activity, 5b was selected for further detailed affinity and efficacy assessments.  28 E. coli BL21 (DE3) cells displaying Hsp90α incubated with 1 µM FITC-labeled p53 lead to a high cellular fluorescence indicating dimerization of Hsp90α. The value obtained was set as 0 % inhibition. In contrast, E. coli cells without displaying Hsp90α (control cells) show no cellular fluorescence. The value obtained here was set as 100 % inhibition. Preincubation of E. coli cells with surface-displayed Hsp90α with 50 µM of the respective substance leads to a lowered cellular fluorescence intensity indicating a lowered binding affinity of FITC-labeled p53 to surface-displayed Hsp90α. These values were set in relation to obtain the relative inhibition of dimerization. c) Apparent K D values of purified CTD of Hsp90α and the respective substance measured via the MST method. A constant amount of 50 nM labeled CTD of Hsp90 was used, and three independent measurements were performed. The resulting mean values were determined and used in the K D Fit formula. d) Cellular viability assessment of a leukemic cell line (K562) measured by incubating with the indicated inhibitors for 72 h, followed by viability measurement using ATP-based Celltitre Glo assay. e) Selection of 5b as a lead candidate on the basis high inhibition of Hsp90α dimerization, low apparent K D, and low IC 50 (µM) in a tested leukemic cell line.

5b binds specifically to CTD of Hsp90α and blocks it cochaperone function
One of the major limitations of NTD-targeting inhibitors is their off-target activity. 4,16 Hence, it is important that the selected hit 5b has a high degree of selectivity against its target, the CTD of Hsp90.
To assess the selectivity of 5b, biochemical cell-free and cellular assays were performed. First, we evaluated the affinity of 5b against Hsp90 in a cell-free assay, where 5b protected recombinant (fulllength) Hsp90ɑ protein in a dose-dependent fashion from degradation against thermolysin enzyme digestion, an assay commonly used to quantify drug affinity-responsive target stability (DARTS) 19,38 ( Figure 3a). Next, we performed the cell-free thermal shift assay 39 Table 1).
NTD targeting Hsp90i Tanespimycin (TM) and PUH-71 served as positive (Hsp90α NTD) or negative control (Hsp90α CTD) in this assay. The thermostabilizing effect of 5b to its target (total Hsp90) was also assessed in a cellular setup, termed cellular thermal shift assay (CETSA) [39][40][41] , a biophysical method based on the ligand-induced thermal stabilization of the protein to directly probe the target engagement in the living cells (SI Table 1, SI Figure 23). The protein quantification for CETSA was performed using a digital western blotter for sensitive and quantitative evaluation of the ligand-protected intracellular Hsp90, whereas TM and PU-H71 served as controls. Next, the thermal stability of intracellular Hsp90 in an increasing concentration of 5b (at a fixed temperature) was determined, a method termed isothermal dose-response fingerprint ITDRF CETSA . 40 5b induced thermal stability of Hsp90 in a dose-dependent fashion, confirming its intracellular and specific target engagement ( Figure   3c, SI Table 1).
Next, to assess the ability of 5b to inhibit Hsp90 chaperone function, a cell-free luciferase-refolding assay [42][43][44] was performed using rabbit reticulocyte lysates as a source of Hsp90. Exposure of 5b decreased the luciferase refolding capacity in a dose-dependent manner by blocking the chaperone function of Hsp90 (Figure 3d). The known Hsp90 NTD inhibitors geldanamycin (GM) and TM served as positive controls. Besides, to assess the specific effect of 5b in obstructing Hsp90 CTD-interacting cochaperones, a time-resolved fluorescence resonance energy transfer (TR-FRET) assay was conducted. 45 5b blocked the binding of PPID (or cyclophillin D, an Hsp90 CTD-interacting chaperone) to recombinant Hsp90α or Hsp90β CTD protein comparable to CA1 treatment, whereas PU-H71, TM, and GM served as negative controls (Figure 3e, Table 1). To rule out the possible interaction of 5b with the NTD of Hsp90α, a fluorescence polarization (FP) competitive assay was carried out using FITClabelled GM 42 (Figure 3f, SI Table 1). As expected, 5b did not show any interaction with the NTD of Hsp90, whereas unlabeled Hsp90 NTD targeting inhibitors GM, Ganetespib (GP), TM, and PU-H71 served as positive controls. Hsp90α (1 µg) was incubated with 5b at indicated concentrations, followed by digestion with thermolysin. Treated protein samples were electrophoresed (SDS-PAGE) and immunoblotted with anti-Hsp90α for detecting protection of Hsp90α protein by 5b (the upper band is protected from proteolysis). b) Cell-free thermal shift assay was performed by incubating recombinant Hsp90α CTD protein with 5b, at an increasing temperature (up to 95 °C). Melting temperature (T m ) without inhibitors (DMSO) was used as a control. c) The dose-dependent intracellular (K562 cells) thermal stabilization (CETSA ITDRF ) of Hsp90 after 5b incubation (24 h) at its increasing concentration (1.25 µM -5 µM). d) 5b inhibits Hsp90α chaperone function, comparable to TM and GM, in cell-free luciferase refolding assay, where the incubation of the inhibitors prevented the rabbit reticulocyte lysate (a source of Hsp90) assisted refolding of denatured luciferase. e) Incubation of 5b blocked the binding of Hsp90 CTD-interacting cochaperone (PPID) in TR-FRET measurements. f) 5b did not reduce the amount of Hsp90-bound FITC-labelled GM and, therefore, does not compete for the GM binding pocket of full-length Hsp90α. Unlabelled GM, GP, PUH71, and TM served as positive controls and NB and CA1 as negative controls.

Binding mode prediction of 5b at Hsp90α
To provide structural insights how 5b binds to the CTD of human Hsp90, we performed 40 independent molecular dynamics (MD) simulations of free diffusion of 5b in the presence of truncated monomeric Hsp90α (aa 294-699), using the Amber 18 suite of molecular simulation programs 46 and the ff14SB 47 and a modified GAFF 30,48 force field for protein and ligand. Initially, we generated 40 individual starting configurations by randomly placing 5b and the CTD structure, leaving at least 10 Å between atoms in 5b and the CTD structure. After minimization, thermalization, and density adaptation, we performed MD simulations of 500 ns length, in which the 5b molecule diffused freely. To counter the high flexibility of the C-terminal helix interface, we introduced positional restraints on the backbone atoms, adjusting the reference coordinates every 100 ns to allow for moderate protein movements.
From the trajectories, first, we extracted all frames where 5b is bound to Hsp90 (no-fit RMSD of 5b ≤ replicas. In the latter case, an area of high density with the shape of 5b is observed (Figure 4a, light blue), which resulted from a single trajectory. This indicates that the ligand was kinetically trapped in this one case, although the position is thermodynamically not favorable. By contrast, the densities in the C-terminal interface are more ambiguously shaped, indicating that, while binding there is favorable, the ligand can still explore multiple binding modes, which are also seen to interchange.
To further study these binding modes, we clustered the bound frames of 5b mapped on the protein surface with respect to their RMSD after superimposing Hsp90. Among the binding modes were several that form interactions to the C-terminal helix interface, with 5b positioned such that it mimics interactions formed by H5' in the dimer (Figure 4c and d). To corroborate that this binding mode is favorable, we computed the effective binding free energies by the MM-GB/SA approach for the trajectory that led to it. Already first transient interactions with the protein resulted in effective energies -12 -down to ~-30 kcal mol -1 . The effective energies decreased further to ~-45 kcal mol -1 once the ligand bound in the C-terminal interface, thereby forming interactions with Hsp90 that remained stable even when the trajectory was extended to 1 µs, indicating that such poses are particularly favorable ( Figure   4b). As to the magnitude of the effective energies, note that configurational entropy contributions were not considered, since estimating such contributions by normal mode analysis may introduce additional uncertainties. 49,50 Overall, the probability density of bound 5b poses, the proportion of replicas, and the results of the MM-GB/SA computations indicate that 5b preferentially binds to the C-terminal helix interface, where it can adopt poses that mimic H5'.

Comparison to 5b binding at Hsp90β
We then set out to study if there is an isoform specificity for the binding of 5b because the helical interface regions differ in three positions: α: S641, β: P633; α: S658, β: A650; α: A685, β: S677 (SI Figure 21). Using the same setup as before, we performed MD simulations of free ligand diffusion around Hsp90β. The probability density of bound 5b again revealed that the C-terminal helix interface is the most preferred region, followed by the cleft between the CTD and middle domain (SI Figure 24).
Notably, no high density in this cleft was found now, in contrast to Hsp90α, confirming that the observation there resulted from kinetic trapping. Hence, despite the few sequence variations in the Cterminal helix interface between Hsp90α and Hsp90β, the same preferred binding region of 5b was found in both cases. a) The relative densities of the bound poses of 5b after 500 ns are mapped on the Hsp90α monomer fragment used in the simulations (PDB ID 3q6m). The missing NTD is shown in red, based on the Hsp90β structure (PDB ID 5fwk). Particularly high densities are observed in the region between H4 and H5 (green circle). A second, less preferred site is in the cleft between the CTD and middle domain (blue circles). b) Effective binding energy calculations over a single trajectory that resulted in 5b binding in the C-terminal helix interface as a function of the center-of-mass distance between 5b and H4 and the simulation time (see color scale). The dashed line at 12.1 Å corresponds to the H4-H5' distance in the crystal structure of PDB ID 3q6m. c) Possible binding mode of 5b in the helix interface, where 5b mimics H5'. d) Blow-up of the possible binding mode of 5b showing how its side chains mimic side chains of H5'.

5b interferes with Hsp90α CTD dimers and disrupts intracellular Hsp90 multiprotein complexes
CTD dimerization of Hsp90 is necessary for its function. 51, 52 To study the effect of 5b exposure on the dissociation of Hsp90 dimers in a cell-free assay, we used Hsp90α CTD protein after incubation with amine-reactive crosslinker BS 3 , as previously described. 42,45 A dose-dependent reduction of Hsp90α CTD dimers along with an increase in the high order oligomeric species was noticed upon incubation with 5b (Figure 5a). Next, we performed small-angle X-ray scattering (SAXS) with the Hsp90α CTD protein, which was coupled to a size exclusion chromatography column (SEC-SAXS) at the ESRF beamline BM29 in Grenoble. 53,54 In the absence of 5b, a clear dimeric profile of the Hsp90α CTD protein was visible on the chromatogram, with an additionally minor tetrameric species (SI Figure 25). We used the program CHROMIXS 55 to merge the frames containing the dimer from this SEC-SAXS profile.
Buffer frames were then subtracted using PRIMUS. 56 From the SAXS data, a radius of gyration (Rg) of 3.23 nm was calculated, which describes the average particle dimension in solution. The ab initio model fit from DAMMIF 57 shows a χ 2 of 1.127, indicating good agreement with the experimental data (Figure 5b and SI Table 2). The corresponding dimeric envelope is highlighted in Figure 5c, superimposed with the calculated dimeric model of Hsp90α CTD. Further, we tested the effect of 5b on the Hsp90 CTD dimer using SAXS (Figure 5d). Due to the low solubility of 5b, we needed to measure the sample as an ensemble of species in solution. First, we tested Hsp90α CTD without 5b on Xeuss 2.0 with Q-Xoom system and observed an increase of Rg to 3.40 nm. This increase is likely due to the small amount of tetramer in solution in the ensemble. Theoretically, the Rg of a tetrameric Hsp90α CTD species is 3.78 nm, using CRYSOL, indicating that even in batch mode SAXS measurements, the Hsp90α CTD protein is predominantly in a dimeric state. We added 5b with an equimolar concentration to Hsp90α CTD protein, and the Rg value slightly increased from 3.40 nm to 3.46 nm. However, with an increasing concentration of 5b to 1 mM, the Rg value increased to 4.11 nm. Compared to the theoretical Rg value of 3.78 nm for the tetramer, we observed that 5b induces oligomerization of Hsp90α CTD to species even larger than the tetrameric form (Figure 5e). It is unclear, however, if the oligomers are formed from Hsp90α CTD monomers or dimers, as the single species could not be resolved in the ensemble measurement.
In a cellular context, Hsp90 acts in multiprotein complexes. 19 Therefore, interfering with Hsp90 function may lead to the disruption of these complexes. In a cellular assay, western blotting was performed under reducing (+dithiothreitol or +DTT) and non-reducing (-DTT) conditions after 5b incubation of the K562 cells. 28 Similarly to AX, 28 5b inhibited the formation of Hsp90 higher-order multimeric species, in contrast to novobiocin (NB), but in concordance with AUY922 (SI Figure 26). Next, to study the effect of 5b exposure on Hsp90 native multiprotein complexes, blue native (BN) PAGE analysis was carried out with K562 cell lysates after 5b incubation. 19,28 At cytotoxic concentrations of 5b, Hsp90ɑ and Hsp90β multiprotein complexes were disrupted, including monomers/dimers of Hsp40 and Hsp27; Hsp60 (primarily in mitochondria) multiprotein complexes, serving as a loading control, were not affected (Figure 5e).
Taken together, these results confirm that 5b interferes with the Hsp90α CTD dimerization, induces oligomerization, and disrupts intracellular Hsp90 multiprotein complexes.

Basic physicochemical properties and microsomal stability of 5b
Next, we assessed the aqueous solubility, chemical stability, and in vitro metabolic stability of 5b. The To study the chemical stability of 5b at physiological pH, the compound was dissolved in a mixture of Tween20/ethanol/phosphate buffer pH 7.5 (7/3/90) and monitored over 24 h (Supplementary Note 2).
Metabolic stability screening of 5b in human liver microsomes revealed 91 % stability after a 40 min incubation at 37 C (Supplementary Note 2). Propanolol, a reference drug with medium to high metabolic stability, showed 74 % of the parent compound remaining, and therefore demonstrated slightly decreased stability compared with 5b. 13 Calculated results for the intrinsic clearance suggest that 5b (6 µL/min/mg) is a low-clearance compound with estimated long half-life (n = 2).

Figure 5: Effect of 5b on Hsp90 oligomeric species and CTD-mediated dimerization. a) Recombinant Hsp90α
CTD was incubated with 63 µM BS 3 crosslinker, with (at indicated concentration) or without 5b, followed by immunoblotting with anti-Hsp90 (AC88) antibody. b) Scattering data of Hsp90α CTD is shown in black dots, with grey error bars. The ab-initio DAMMIF model fit is shown as a red line. The intensity is displayed as a function of momentum transfer s. c) The volumetric envelope, calculated from the scattering data using DAMMIF 57 , is shown as a blue surface. The monomers of the predicted Hsp90 CTD dimer model are shown in green and cyan. Superimposing was done using SUPCOMB. 57 d) The Radius of gyration (Rg) of the different Hsp90α CTD protein samples was calculated using the Guinier approximation. 58 The theoretical Rg of the tetramer was calculated using CRYSOL based on the structure PDB ID 3q6m. 59 e) Native Hsp90 complexes in K-562 (24h administration of 5b) was identified by running Blue Native (BN) gels followed by immunoblotting analysis. The cytotoxic concentration of 5b resulted in the potent disruption of Hsp90ɑ, Hsp90β, Hsp40, and Hsp27 complexes and monomers/dimers. AUY922 exposure elevated the expression of HSR associated protein complexes and monomers/dimers (Hsp40 and Hsp27), whereas Hsp60 served as loading controls.

5b is effective against resistant leukemia cells and in the zebrafish xenotransplantation model
Elevated Hsp90 expression is reported in several resistant leukemia subtypes such as BCR-ABL1 + CML/BCP-ALL, FLT3-ITD-driven AML, and Ph-like BCP-ALL. [60][61][62][63] Besides that, client proteins of Hsp90 include several kinases (e.g., AKT, BCR-ABL1, BRAF, EGFR2, HER2, and JAK1/2), growth and steroid receptors, and apoptotic factors (e.g., BCL-2 and mutant p53), which are often aberrantly regulated in several malignancies. 2,64 Therefore, we determined the efficacy of 5b on therapy-resistant cell lines obtained from different leukemia subtypes (B-ALL, T-ALL, CML, and AML), including imatinib (IM), multi TKI-resistant and bortezomib (BTZ)-resistant leukemic cell lines. 28 Average IC 50 values of 5b in these cell lines were determined using ATP-based viability assay and plotted in a clustered heat map, indicating the superior efficacy against BCR-ABL1+ and AML leukemic cell lines, when compared to T-ALL cell lines (Figure 6a).
As previously performed with AX 28 , we next determined the destabilizing effect of 5b on BCR-ABL1 expression and associated downstream signaling. In K562 cells, 48 h exposure of 5b down-regulated the phospho-BCR-ABL1 and total-BCR-ABL1 levels, as well as the related downstream signaling pathways, as evaluated by immunoblot analysis (Figure 6b). 5b additionally reduced the expression of client proteins associated with Hsp90 chaperone activity, involving Akt, Stat5, and c-Myc (Figure 6b).
In contrast to AUY922, the exposure of 5b on the K562 cells did not induce the expression of Hsp70, Hsp40, and Hsp27 involved in HSR (Figure 6c). Exposure of 5b to the leukemic cell lines (K562, KCL22 and HL60) inhibited their proliferation (SI Figure 27) and induced apoptosis in a caspase 3/7 enzyme-dependent assay, with induction of approximately two-to eight-fold increase of apoptotic cells, in accordance to the reference Hsp90 CTD-targeting inhibitor, novobiocin (NB) (SI Figure 28).
Besides, 5b facilitated early differentiation measured by expression of differentiation markers involving CD14 vs. CD11b in HL60 cells, and CD133 vs. CD11b in Mutz-2 cells (SI Figure 29). In this line, 48 h exposure of 5b to K562 cells significantly reduced the colony-forming capacity (SI Figure 30). To further evaluate the efficacy of 5b on leukemic cells, we used the zebrafish xenotransplantation model 65 (Figure 6d). MOLT-4 cells were transplanted into zebrafish embryos at 32 hours post-fertilization (hpf).
At this stage, the adaptive immune system is not yet developed, therefore, human cells can be tolerated by the host. The transparency of zebrafish embryos also enabled us to monitor the distribution of human cells, which were stained with a vital fluorescent dye. Xenotransplanted embryos were treated with AUY922 (500 nM) and 5b (500 nM) for 48 h, whereas DMSO was used as a negative control. No toxicity of drugs at the given concentration was noticed on the development of xenotransplanted embryos. We then determined the number of MOLT4 cells in each treated group using flow cytometry.
Compared to the DMSO-treated group, the number of transplanted cells was significantly reduced in AUY922-and 5b-treated embryos.
To conclude, this data confirms the anti-leukemic effect of 5b without inducing HSR.  50 data was plotted as a clustered heat map, followed by unsupervised hierarchical clustering. The vertical axis of the dendrogram exemplifies the dissimilarity between clusters, whereas the color of the individual cell is related to its position along a log IC 50 (µM) gradient. b) Treatment of K562 cells with 5b and respective controls (AUY922 and NB) for 48 h resulted in the downregulation of BCR-ABL1+ and subsequently downstream signaling pathways including phosphorylated and unphosphorylated Stat5a, Crkl, Akt, S6 (mTOR), and cMyc. c) K562 cells were treated with the indicated (cytotoxic) concentration of 5b, NB, and AUY922 for 48 h, and later protein lysates were subjected to immunoblot analysis. As expected, 5b and NB did not induce expression of Hsp70, Hsp40, and Hsp27, whereas AUY922 leads to HSR induction. Hsp60 (primarily present in mitochondria) and PDI (endoplasmic reticulum) served as a control.

5b acts on TKI-resistant BCR-ABL1+ leukemic cells
The integration of specific tyrosine kinase inhibitors (TKI) such as imatinib (Gleevec) into polychemotherapy treatment protocols has significantly improved the response rate in BCR-ABL1+ leukemia patients (initial remission went from 35% to 88%). 66 However, stable remission cannot be sustained in many cases as the leukemic cells pursue several escape mechanisms against TKI treatment; one of them is the occurrence of mutations in the ABL1 kinase domain. Especially, in the BCR-ABL1 T315I mutant, only ponatinib (TKI) is effective 67 , albeit with severe cardiovascular side effects. 68 As Hsp90 facilitates the correct folding of several oncogenic newly synthesized or denatured proteins, among them BCR-ABL1, 69-71 therefore targeting Hsp90 with small molecule inhibitors would destabilize BCR-ABL1 and can serve as a therapeutic target. 3, 10 We, therefore, next tested the efficiency of 5b in a murine BA/F3 cell line stably expressing clinically relevant mutants (BCR-ABL1 T315I , BCR-ABL1 E255K , and BCR-ABL1 M351T ) with prominent TKI resistance profiles. 28 As expected, similar to K562 cells, exposure of 5b significantly reduced proliferation (Figure 7a) and induced apoptosis (Figure 7b) at 6 and 12 µM, comparable to NB (at 0.3 mM) in BA/F3 cells expressing BCR-ABL1 T315I, E255K, or M351T mutants. Additionally, after the application of 5b, BCR-ABL1 oncoprotein was destabilized, and downstream signaling pathways (Akt and Stat5) were blocked with increasing concentrations of 5b, comparable to the human leukemic cell lines (Figure 7c). Furthermore, 24 h exposure of 5b on BCR-ABL1 T315I, E255K or M351T mutants-expressing BA/F3 cells significantly inhibited the colony formation ability (Figure 7d). In addition, in our generated human BCR-ABL1+ IM-resistant cell line models (K562-IMr and KCL22-IMr), 28 5b did not differentiate in inducing apoptosis between IM-resistant vs. IM-sensitive clones, proposing a superior effect of 5b in human IM-resistant BCR-ABL1-positive cells (Figure 7e and SI Figure 31). Encouraged by these results, we next tested 5b on three primary CML CD34+ IM-resistant patient samples in the range of cytotoxic concentration (5-10 µM) obtained from leukemic cell lines. Similar to BCR-ABL1+ leukemic cell lines, the exposure of 5b reduced the proliferation, induced apoptosis, and reduced the colony-forming ability of CML CD34+ IMresistant patient cells and also revealed a potent combinatorial inhibitory effect when used in combination with IM (Figure 7f -7h). assay. c) Exposure of 5b to these cells destabilizes BCR-ABL1 and subsequently its associated downstream signaling pathways. d) BA/F3 BCR-ABL1 mutant-expressing cells were seeded in methylcellulose-based semisolid medium, after 24 h treatment with 5b. Colonies were counted after 14 days. e) K562-IMr cells were treated with the indicated concentration of 5b for 48 h, later dually stained with annexin V/PI, and subsequently measured by FACS. f) Primary CML CD34+ patient cells were treated with the indicated concentration of 5b or IM alone or in combination, and later the viable cells were counted after every 24 h interval for 6 days. g) Apoptosis induction in primary CML CD34+ patient cells after exposure of 5b or IB alone or in combination. g) Primary CML CD34+ patient cells were seeded in methylcellulose medium after 24 h treatment with 5b, IM alone or both in combination. Colonies were counted after 14 days. Columns depict the mean of three independent experiments (n = 3).

Discussion
The CTD of Hsp90 contains several binding areas: the C-terminal ATP binding site, the MEEVD motif at the end of the CTD, and the primary dimerization interface of Hsp90. 23 The C-terminal ATP binding site is only available upon occupation of the N-terminal ATP binding pocket and binds purine and pyrimidine nucleotides, while the N-terminal ATP binding site (NTD) is more specific for adenine derivatives. 72 The MEEVD motif binds the TRP-domain of cochaperones such as Hop and immunophilins, which is formed by a four-helix bundle and is crucial for Hsp90 dimerization. 2 Proteinprotein interactions in the interface of the four-helix bundle maintain the dimeric state of Hsp90. 51 In this study, we present the development of the first-in-class small-molecule inhibitor of Hsp90, 5b, which was rationally designed to target the C-terminal dimerization interface. Based on our prior hotspot prediction 26 and the prototype compound AX 28 , we successfully performed a scaffold-hopping from the aminoxy peptide-backbone towards more drug-like tripyrimidones. 5b mimics α-helical side-chains that form hotspot residues located on H5 in the dimerization interface. To independently predict the binding site and mode of 5b, we performed extensive MD simulations, in which the inhibitor was allowed to diffuse freely around an MD-CTD construct of Hsp90α, similar to our analysis on AX binding 28 and related studies. [73][74][75] The results revealed the C-terminal dimerization interface as the most likely binding region of 5b, which was confirmed by effective binding energy computations, corroborating the hypothesis underlying the design of 5b. Following the high sequence-identity in the interface region, similar results were obtained for Hsp90β, suggesting that 5b does not exhibit isoform specificity.
Next, the selective binding of 5b to Hsp90 was validated in a variety of biochemical and cellular assays, including DARTS, thermal and isothermal shift assays, whereas intracellular Hsp90 engagement and disruption of Hsp90 multiprotein complexes were established via CETSA and ITDRF CETSA assays, and immunoblotting under reducing conditions (+/-DTT) and non-denaturing (BN-PAGE) conditions. Moreover, a crosslinker and auto display dimerization assay, as well as SEC-SAXS measurements, repeatedly confirmed the destabilization of Hsp90α CTD dimers upon 5b incubation, whereas no unspecific binding of 5b was reported on the Hsp90α NTD protein in an FP-based competitive assay.
However, during SAXS measurements, which were performed on an ensemble of species in solution, we were unable to determine whether 5b-induced oligomers originated from Hsp90α CTD monomeric or dimeric species. In addition, 5b blocks the chaperone function of Hsp90, as determined by TR-FRET assay and in cell-free luciferase refolding assay. Markedly, even though there are differences in the inhibitory concentrations against tested leukemia cells between 5b (IC 50 in submicromolar range) and reference Hsp90 NTD control inhibitors (IC 50 in subnanomolar range), a comparative selectivity profile (in cell-free or cell-based biochemical assays) toward Hsp90 was observed between 5b and Hsp90 NTD reference inhibitors. This data indicate that the conventional Hsp90 NTD inhibitors induce cellular inhibitory effects through their off-target activity, besides targeting Hsp90. 4,16 Moreover, 5b exhibited potent in vitro anticancer activity against a broad spectrum of therapy-resistant leukemia cell lines (including TKI and proteasome inhibitor-resistant) and primary TKI-resistant (BCR-ABL1+) leukemia patient cells. 5b significantly reduced the leukemia burden in the zebrafish xenotransplantation model and induced apoptosis in TKI-resistant BCR-ABL1 T315I, E255K or M351T mutant cells by destabilizing the BCR-ABL1 expression and, thereby, hampering related downstream signaling cascades without HSR induction. These data collectively established 5b as a first-in-class smallmolecule inhibitor that targets the C-terminal dimerization interface.

Conclusion
Through structure-based molecular design, chemical synthesis, molecular simulations-based prediction of the binding mode, and evaluation of biochemical affinity, we have developed the first low-molecular weight compound interfering with the Hsp90 CTD dimerization. The C-terminal Hsp90 inhibitor 5b contains a tripyrimidonamide scaffold and is active against therapy-resistant leukemia cells as well as in a zebrafish xenotransplantation model without exhibiting the pro-survival resistance mechanism HSR.

Methods
Chemical synthesis: See Supplementary Note 1 for general methods, synthetic protocols, compound characterization, and spectral data (SI Figure 1 -20).

Molecular dynamics (MD) simulations:
The structures of Hsp90α and β (PDB IDs 3q6m and 5fwk respectively) were prepared using Schrödinger Maestro 80 . For each of the isoforms, 40 individual MD simulations were performed. Initial random placement of 5b and solvation in TIP3P water 81 was done using PACKMOL 82 , neutralizing the system by the addition of sodium ions. ff14SB was used as force field for the protein and a modified GAFF version 1.5 for 5b. 26,48 All simulations were carried out using the Amber18 software package. 46 To treat long-range electrostatics, the particle mesh Ewald method 83 was used with a cutoff of 9.0 Å for equilibration and 10.0 Å for production. The SHAKE algorithm 84 and hydrogen mass repartitioning 85 were used to allow for simulation steps of 2 fs in the equilibration and 4 fs in the production.
Initially, the systems were energy-minimized using the steepest descent (500 steps) and conjugate gradient (2000 steps) methods and placing positional restraints with a force constant of 5 kcal mol -1 Å -2 on all protein atoms; the restraints were reduced in a second energy minimization to a force constant of 1 kcal mol -1 Å -2 (for 2000 steps of steepest descent followed by 8000 steps of conjugate gradient), and removed in a third one (for 1000 steps of steepest descent followed by 4000 steps of conjugate gradient).
Placing positional restraints with a force constant of 1 kcal mol -1 Å -2 on the backbone atoms, first, the system was heated to 100 K in 50 ps of NVT MD and further heated to 300 K in 5 ps of NPT MD.

MM-GB/SA computations:
Effective binding energies were computed over one MD trajectory that resulted in binding of 5b in the CTD dimerization interface and led to a binding mode in which 5b mimics H5'. In the computations, the single trajectory approach was used, where complex, protein, and ligand configurations were extracted from the complex trajectory. 87 After removing water molecules and counterions, gas-phase energies (van der Waals and electrostatic contributions) were evaluated on every frame sampled at an interval of 200 ps using MMPBSA.py 88 ; the polar contribution to the solvation free energy was calculated using the "OBC II" generalized Born model 89  Signaling Technology and anti-β-actin (Sigma-Aldrich). Blue-native (BN) gels were performed following manufacturer's instructions (Invitrogen) and performed previously. 28 Briefly, lysates were generated from K562 cell line after 48 h treatment with inhibitors (at indicated concentration) using NativePAGE Sample Prep kit (Invitrogen) by 2-3 freezing thawing cycles followed by centrifugation at 20,000 x g for 25-30 min at 4°C.
Dimerization assay: Hsp90 CTD dimerization was evaluated using an amine-reactive chemical crosslinker bis(sulfosuccinimidyl) suberate (BS 3 ) (Pierce). 42,45 Hsp90α CTD protein (2 µM) was diluted in Na 2 HPO 4 (25 mM; pH 7.4) and treated with different concentrations of the inhibitor to make a final volume of 25 µL. The reaction mixture was incubated at RT for 1 h. The amine-reactive crosslinker BS 3 was added to a final concentration of 63 µM and the samples were incubated for 1 h at RT. Crosslinking was quenched by the addition of SDS sample buffer and subsequent heating for 5 min at 95 °C. Samples were run in 12% SDS-PAGE gels followed by western blotting. Blots were probed with anti-Hsp90 (AC88, Abcam) antibody.

SEC-SAXS:
We collected the SEC-SAXS data on beamline BM29 at the ESRF Grenoble 53,54 . The BM29 beamline was equipped with a PILATUS 2M detector (Dectris) at a fixed distance of 2.827 m.
The measurement of Hsp90 CTD (18 mg/ml) was performed at 20°C on a Superdex 200 increase 3.2/300 column (Buffer 50 mM TRIS pH 7.5, 100 mM NaCl) with a flowrate of 0.075 ml/min, collecting one frame each two seconds. Data were scaled to absolute intensity against water. Further, we have collected SAXS data on our Xeuss 2.0 Q-Xoom sytem from Xenocs, equipped with a PILATUS 3 R 300K detector (Dectris) and a GENIX 3D CU Ultra Low Divergence x-ray beam delivery system. The chosen sample to detector distance for the experiment was 0.55 m, results in an achievable q-range of 0.10 -6 nm -1 . All measurements were performed at 20°C with protein concentrations of 9.7 and 10.8 mg/mL. Compound 5b was added and incubated for 30 min at 20 °C. Samples were injected in the Low Noise Flow Cell (Xenocs) via autosampler. For each sample, 18 frames with an exposer time of ten minutes were collected. Data were scaled to absolute intensity against water. All used programs for data processing were part of the ATSAS Software package (Version 3.0.3) 90 . Primary data reduction was performed with the programs CHROMIXS and PRIMUS 55,56 . With the Guinier approximation 58 , we determine the forward scattering I(0) and the radius of gyration (R g ). The program GNOM 91 was used to estimate the maximum particle dimension (D max ) with the pair-distribution function p(r). Low resolution ab initio models were calculated with DAMMIF 57 . Superimposing of the predicted model was done with the program SUPCOMB 92 .
Physicochemical properties of 5b (See Supplementary Note 2 for more details):

Aqueous solubility of 5b
The aqueous thermodynamic solubility of 5b was determined in phosphate-buffered-saline (PBS, pH 7.4) after 4 and 24 h incubation time at 25°C. Ondansetron was used as reference compound with high solubility of 95 µM. The thermodynamic solubility of 5b was ranging from 4 µM after 4 h to 8 µM after 24 h (n = 2). For detailed information see Bienta, Enamie Biological Services study reports.

Chemical stability of 5b
Drug decomposition was determined by high-performance liquid chromatography (HPLC, Method 1).
The stability of 5b at acidic pH was determined by dissolving 5b in a mixture of Tween20/ethanol/phosphate buffer pH 2 (7/3/90) and the stability monitored over a period of 24 h at 37°C. After 24h, only slight decomposition was detected (1.3 % drug decomposition, n = 2).

In vitro metabolic stability of 5b in human liver microsomes
The metabolic stability screening of 5b in human liver microsomes revealed 91 % stability after a 40 min incubation at 37°C. Propanolol, a reference drug with medium to high metabolic stability showed 74 % of the parent compound remaining, and therefore demonstrated slightly decreased stability compared with 5b. Calculated results for the intrinsic clearance suggest that 5b (6 µL/min/mg) is a low clearance compound with an estimated long half-life (n=2). For detailed information see Bienta, Enamie Biological Services study reports. supplemented with or without IL-3 (10ng/ml) respectively. BA/F3 cells expressing BCR-ABL1 mutants were resistant against imatinib (IM) until ~10 µM. 28 IM resistant BCR-ABL1+ K562 (K562-IMr), KCL22 (KCL22-IMr) and SUPB15 (SUPB15-IMr) were generated by gradual increase (1-2.5 µM) in the concentration of IM (Sigma-Aldrich, St. Louis, MO, USA) over a period of 3 months. 28,93 Bortezomib (BTZ) resistant clones (80 nM) of HL60 (HL60-BTZr) was established following similar protocol as described to pick IM resistant clones. Primary patient derived CML CD34+ blast cells were cultured in mononuclear cell medium (PromoCell, Heidelberg, Germany).
Viability assay: Cells were seeded in white 96-well plate (Corning, NY, USA) with increasing concentration (50 nM -25 µM) of inhibitors and respective controls for 48 h. Cell viability was monitored using Celltitre Glo luminescent assay (based on the ATP quantification), following manufacturer's guidelines (Promega). 28 IC 50 for compounds were determined by plotting raw data (normalized to controls) using sigmoid dose curve and nonlinear regression (GraphPad Prism). Xenotransplantation in zebrafish embryos: Xenotransplantation experiment was performed as described previously. 65 Briefly, MOLT-4 cells were labeled with Vybrant™ CFDA SE Cell Tracer Kit (Invitrogen) following the manufacturer's instructions, and then were suspended in PBS at a density of 1 × 10 8 cells/ml. Approximately 1 nl cell suspension (around 200 cells) was injected into the perivitelline space of embryos at 32 hpf. Injected embryos were first incubated at 28°C for one hour. Only embryos with good engraftment were selected for treatment with DMSO (control group), AUY922 (500 nM) or 5b (500 nM) for 48 hours at 35°C. This temperature enables the maintenance of embryos with grafted cells without compromising zebrafish development. Drug-treated embryos were dissociated by passing through a 40-µm cell strainer (Greiner Bio-One) and then analyzed using a BD LSR II flow cytometer.
Fold change of engrafted MOLT4 cells was calculated to mean of DMSO-treated embryos. GraphPad Prism software (version 7) was used for graphing and statistical analysis.

Notes:
The authors declare no competing financial interest.