Exploiting Ligand-binding Domain Dimerization for Development of Novel Androgen Receptor Inhibitors

Recurring prostate cancer in its late metastatic stage is traditionally treated by inhibition of the androgen receptor (AR) via androgen deprivation and AR antagonists (1). Several AR antagonists such as enzalutamide (Enz) and apalutamide are currently being tested in earlier stages, even as neoadjuvant treatment before surgery of localized, high-risk prostate cancer (2). The AR is a ligand-activated transcription factor that responds to testosterone (T) and dihydrotestosterone (DHT) for supporting the development of male reproductive organs and fertility (3). Next to an intrinsically disordered amino-terminal domain, the AR contains a DNA-binding domain and a ligand-binding domain (LBD) that contains a helical fold, conserved among nuclear receptors (4). Several functions coincide within the LBD. A hydrophobic ligand-binding pocket (LBP) is buried within the LBD, while the surface contains three well-described protein–protein interaction surfaces: the coactivator binding groove (AF2), binding function 3 (BF3), and the LBD dimerization interface (4, 5).

Despite the initial success of AR signaling inhibitors when given in the metastatic setting of prostate cancer, resistance occurs after several months to years. Most resistance mechanisms involve changes that reactivate the full-size AR either by AR upregulation, by changing ligand specificity through point mutations like T877A or by changes in the coregulator balance; while in some cases alternative signaling pathways are activated (6, 7). In addition, AR-targeted therapy can lead to the expression of AR splice variants (ARV) which might play a role in resistance (8). In vitro, these splice variants such as AR-V7 are constitutively active because they lack the LBD. In clinical samples, they correlate with overexpression of full-length AR (AR fl). Because of heterodimer formation between AR fl and its splice variants, even disease that expresses these variants retains ligand control (9).

Because most resistance mechanisms still involve the AR, research groups have focused on other domains of the AR or on other surfaces within the LBD to keep AR activity in control. EPI-7170 for example demonstrates synergistic effects with Enz and is effective against ARVs by inhibiting the AR via the NTD (10). Pharmacokinetic optimization of a BF3 inhibitor series resulted in the VPC-13822 prodrug that inhibits AR interactions with coregulators and as a result also reduces tumor volume in animal models of castration-resistant prostate cancer (11). Up until now, the LBD dimerization surface has remained unexplored as a target for AR inhibition. The LBD dimerization surface of about 1,000 Ų is centered around the end of Helix 5 of both LBDs. Additional contributions to the interface are mediated by residues in H1, H7-H9, L1-L3, and S1 (5). The presence of several disease-causing mutations within this surface already suggested its necessity for proper functioning of the AR (5). The functional relevance of LBD dimerization has been demonstrated in the AR^Lmon mouse model where LBD dimerization is disrupted by the W731R mutation, located at the end of Helix 5 (12). AR^Lmon/Y mice display a feminized external and internal phenotype with nipple formation, small anogenital distance, underdeveloped interabdominal testes, and a complete lack of male accessory sex organs. Mechanistically, this can be explained by the inability of AR W731R to recognize specific AR-binding sites in the DNA, albeit the receptor was shown to enter the nucleus and occupy chromatin in AR^Lmon/Y kidneys (12). On the basis of this mouse model, we conclude that LBD dimerization is crucial for normal physiologic AR activity, making this interface an excellent candidate for therapeutic targeting in prostate cancer.

We here report the discovery of AR inhibitors that disrupt LBD dimerization via an in silico screening and validate their noncompetitive nature compared with Enz and DHT. Moreover, we demonstrate their antiproliferative activity on AR-positive prostate cancer cell lines and synergistic activity in concert with Enz.

The crystal structure of the human AR LBD homodimer [PBD: 5jjm (5)] and its monomer were prepared using three-dimensional (3D) protonation, and energy refinement with MOE v.2018 (RRID:SCR_014882; ref. 13). Databases that contain over 1 million commercially available compounds in two-dimensional formats were prepared using MOE Database Wash where the molecules were altered to their dominant protomer/tautomer at the pH of 7. Ready-to-dock 3D conformation databases were generated with MOE Conformation Import in which a maximum of 1,000 conformations were produced for each molecule. The Gromacs v.2018 package (RRID:SCR_014565; ref. 14) was used to perform molecular dynamics (MD) simulations with the amber99sb force field (15), three-points (TIP3P) water model (16), 1 ns V-rescale thermostat NVT ensemble (17), 1 ns V‐rescale thermostat and Parrinello-Rahman barostat NPT ensemble (18), LINCS algorithm (19), and PME implementation of the Ewald method (20). The binding pockets of the AR LBD monomer were identified using FTmap (21). To filter the 3D conformations databases, 3D pharmacophore and molecular docking–based virtual screening methods were combined via MOE Pharmacophore Search and GOLD v.2018 (ref. 22; RRID:SCR_000188), respectively. In the GOLD package, a genetic algorithm and the Goldscore (23) scoring function were applied to produce 20 docking poses for each molecule. The docking output was further filtered by applying the DSX (24) knowledge-based scoring function to remove any compound with incompatible interactions.

DHT (10300) and 3,3,5-triiodothyroacetic acid (TRIAC; T7650) were purchased from Sigma-Aldrich. Enz (SRP016825m) was purchased from Sequoia Research Products. Dimerisation Inhibiting Molecules (DIM) compounds were purchased from Specs (www.specs.net). Full characterization of DIM20, DIM20.39, DIM24, and DIM25 can be found in Supplementary Fig. S1. All compounds were dissolved in DMSO (Biotium). Dynamic Light Scattering (DLS) using Malvern Zetasizer ZSP confirmed solubility of the compounds.

The ClARE cell line was described before (25). Briefly, the ClARE cell line is a HEK293 Flp-In cell line with a random integrated human AR cassette and the pGL4-luciferase reporter gene under control of an E1B TATA box and four copies of the Slp-HRE 2 mutated (5′ AGAACTggcTGTCCA 3′) in the Flp Recombination Target site. The cell lines LNCaP (RRID:CVCL_0395), PC3 (RRID:CVCL_0035), HEK 293 (DSMZ, catalog no. ACC-305, RRID:CVCL_0045), VCaP (RRID:CVCL_2235), and 22rv1 (RRID:CVCL_1045) were obtained from ATCC and cultured as recommended. Authentication via short tandem repeat profiling was performed by Genetica DNA Laboratories for validation. Mycoplasma test were performed at least once every 6 months. All experiments were performed within 20 passage numbers after thawing the cell lines.

ClARE cells (25) were seeded in 96-well plates at a density of 15,000 cells per well in DMEM-5% CSS. The next day, medium was refreshed and replaced with compound/vehicle dilutions in DMEM-5% CSS. Assay is performed in duplicate with one replicate for metabolic activity and the other to measure luciferase activity.

After 24 hours of incubation with compound, metabolic effects of the compounds were determined with the Cell Proliferation Assay XTT (Applichem) or Cell Counting Kit-8 (Dojindo EU Gmbh) according to instructions of the manufacturer. For luciferase activity, the cells were lysed in Passive lysis buffer (Promega). Luminescence was measured after adding Luciferase Assay Reagent (Promega). Luciferase values are given relative to the DHT 1 nmol/L condition which is set at 100%. Mean and SD of at least three independent datapoints are given. IC₅₀ and CC₅₀ are calculated using nonlinear regression [log(inhibitor) vs. normalized response—variable slope] in Graphpad Prism.

The Nanobit system consisting of LgBit-AR and AR-SmBit (Promega) under the control of the cytomegalovirus promoter was transfected in HEK 293-T cells (KCB, catalog no. KCB 200744YJ, RRID:CVCL_0063) using Genejuice Transfection Reagent (Merck Life Science). Two days later, medium was refreshed with DMEM-5% CSS supplemented with HEPES (10 mmol/L) and the corresponding compound. Cells were pretreated for 30 minutes with the compound, then furimazine (KyvoBio; 8 μmol/L) was added and baseline was measured every minute for a total of 5 minutes. After addition of DHT, luminescence was recorded for 70 minutes in steps of 5 minutes.

The proliferation measurements were performed as described before by live cell analysis (IncuCyte ZOOM, Essen BioScience). For LNCaP, a nuclear mKate2 fluorescent label was used to count nuclei. Nuclei counts were log2 transformed (26). For PC3, HEK 293 and 22rv1 Phase Object Confluence (Percent) was monitored for at least 72 hours. The proliferation rate was measured once the cells maintained a stable growth rate.

Cell viability of VCaP cells was determined via CellTiter-Glo Luminescent Cell Viability Assay (Promega) after 3 and 6 days of incubation with DIM compounds. Each 96-well was seeded with 30,000 cells in DMEM-10%FCS, the next day compounds were added at a final concentration of 3, 10, or 30 μmol/L. All values are given relative to the 3 μmol/L vehicle-treated cells which were set at 100%.

Via a whole-cell competition assay, binding of a ³H-DHT (Perkin Elmer) gradient to the AR is measured in absence or presence of a constant amount of modulator (DIM compound or control). On the first day, 45,000 ClARE cells were seeded per well of a 48-well plate in DMEM-5% CSS. Two days later, the medium was replaced by DMEM-5% CSS containing increasing amounts of ³H-DHT (0.1, 0.3, 1, 3, 10, 30, and 100 nmol/L). The DIM compounds were present at a constant concentration of either 5 or 25 μmol/L while Enz was present at 10 μmol/L. Specific binding of ³H-DHT was obtained by subtracting the nonspecific binding [³H-DHT-gradient + excess of DHT (10 μmol/L)] from the total binding (³H-DHT-gradient).

ClARE cells (25) were seeded in 96-well plates at a density of 15,000 cells per well in DMEM-5% CSS. The next day, 2.5 ng of the expression plasmid for the AR NTD-DBD (1–669) was cotransfected with 5 ng of pCMV B-Gal plasmid in each well using GeneJuice Transfection Reagent (Novagen). Empty vector was used to increase the total transfected DNA up to 100 ng per well. The following day, medium was refreshed and replaced with 1 nmol/L DHT supplemented with compound and/or vehicle in DMEM-5% CSS. On the fourth day, cells were lysed in Passive Lysis Buffer (Promega) and luciferase and B-Gal activity were measured using Luciferase Assay Reagent Buffer (Promega) or Galactolight Buffer, Galacton+ and Light Emission Accelerator (Applied Biosystems). Mean and SD of three independent biological replicates are given.

Combination index was calculated for cotreatment of DIM compound and Enz on the ClARE cell line using the CompuSyn software (27). Drugs were combined at a constant ratio that is near IC_50-DIM/IC_50-Enz. Average effect values (Fa = affected fraction; n = 4) for each drug concentration were used as input of the program. Five concentrations with a 2-fold serial dilution are used for each single drug (0.25×IC₅₀, 0.5×IC₅₀, IC₅₀, 2×IC₅₀, 4×IC₅₀) and the combinations of these are used as two-drug treatment. The IC₅₀ of DIM20 and Enz were determined on beforehand and correspond to 1.30 μmol/L and 0.125 μmol/L, respectively.

For cotreatment of DIM compound and Enz on the LNCaP cell line, drug combinations at a nonconstant ratio were used. The concentration of DIM compound was 5, 10, or 20 μmol/L while the concentration of Enz was 250 nmol/L or 10 μmol/L.

A pET expression plasmid containing Thioredoxine-His-Thrombin-TEV- AR LBD (663–919)- GSGSGS-Avi-tag (GLNDIFEAQKIEWHE) was cotransformed with pSAM-2xflag-BirA in in BL21(DE3)-RILP codon plus competent cells (Agilent Technologies). This pSAM-2xflag-BirA was created by integrating the BirA cassette from pET21d-myc-BirA into pSAM using specific primers that contain a double flag-tag and sufficient overhang to perform infusion cloning. pET21d-myc-BirA was a gift from Martin Spiess (Addgene plasmid # 109424; http://n2t.net/addgene:109424; RRID:Addgene_109424; ref. 28) and pSAM was a gift from Jumi Shin (Addgene plasmid # 45174; http://n2t.net/addgene:45174; RRID:Addgene_45174; ref. 29). Protein expression is induced with 1 mmol/L IPTG and in presence of 20 μmol/L DHT and 125 μmol/L biotin (Sigma-Aldrich) at 15°C for 18 hours. Bacterial pellets were lysed by incubation with lysozyme and subsequent sonication. The AR-LBD with avi-tag was purified by Immobilized Metal Chelate Affinity Chromatography (IMAC), followed by TEV protease digestion, and size exclusion chromatography at 4°C. The final buffer composition was 50 mmol/L HEPES pH7.0, 150 mmol/L Li₂SO₄, 10% glycerol, 1 mmol/L dithiothreitol (DTT), and 20 μmol/L DHT.

For the biolayer interferometry (BLI) experiments, the Octed RED (Fortebio) was used to quantify binding of compounds to the AR-LBD at 30°C. The BLI buffer contains 150 mmol/L Li₂SO₄, 50 mmol/L HEPES pH 7.0, 1 mmol/L DTT, and 500 nmol/L DHT.

Purified biotinylated AR-LBD (0.1 μg/μL) in BLI buffer was bound to super-streptavidin (SSA) sensors overnight at 4°C. Simultaneously, for each compound, a reference sensor was incubated in biocytin (Sigma-Aldrich) 10 g/L in BLI buffer. Both sensors were further blocked with 10 g/L biocytin in Superblock Blocking (TBS) buffer (Thermo Fisher Scientific) for 1 hour at room temperature and further equilibrated in BLI buffer and BLI buffer 1% DMSO. For detection of unspecific binding of the compound, the biocytin-reference sensors were dipped in the same compound solutions and this signal was subtracted from the signal measured with the LBD sensors. A biocytin sensor and a LBD-sensor dipped in buffer was used to account for buffer drift. Binding of the DIM compound or the AR20-30 peptide (RGAFQNLFQSV; GenScript) to AR LBD was measured at different concentrations via consecutive binding cycles (1 minute baseline- 1 minute association- 3minutes dissociation). The K_D for each compound was calculated via steady-state analysis using the K_obs Linearisation Method of the Anabel software (30). The average of two independent measurements was taken.

Log transformation was performed on the dose–response experiments. After the log-transformation of the concentrations, we performed nonlinear regression with GraphPad Prism v9 (GraphPad Software, LLC; RRID:SCR_002798) using the formula “log(agonist) versus response (three parameters)” or “log(agonist) versus response—variable slope (four parameters)” which resulted in LOG-EC₅₀ values with their SEMs and 95% confidence intervals for each treatment.

Statistical analyses were performed using GraphPad Prism v9 (GraphPad Software, LLC; RRID:SCR_002798). Normality and lognormality tests were performed on the continuous data using the D'Agostino–Pearson normality test or the Shapiro–Wilk test in case of small datasets. Datasets that were not sampled from a normally distributed population were further analyzed using nonparametric tests. To compare three groups or more, an ordinary one-way ANOVA with Dunnett multiple comparisons test was performed; the vehicle-treated condition is the control sample. An ordinary two-way ANOVA with Dunnett multiple comparisons test was performed in experiments with more than one independent variable, vehicle-treated condition was used as reference for the multiple comparisons. For the time course experiments, a Friedman test combined with Dunn multiple comparisons test was used with wild-type (WT) as the control. P values or adjusted P values are mentioned within the graphs whenever they reached significance (P values below 0.05). Box-whisker plots depict the spread of the data with the box extending from the 25th to 75th percentiles including the median as a line and the whiskers showing the minimal and maximal value. Dot plots show the individual datapoints with the mean as a horizontal line.

The data generated in this study are available upon request from the corresponding author.

Our structure-based virtual screening consisted of five main stages: (i) AR dimer crystal structure validation, (ii) druggable binding site identification, (iii) pharmacophore search, (iv) molecular docking, and (v) visual evaluation (Fig. 1). The MD simulations showed that the crystal structure of AR LBD dimer [PDB ID: 5jjm (5)] is stable during simulation time, with a minor hinging movement between the two subunits, hence the dimer interface is not a crystal packing artifact and can be used as a platform for structure-based drug design (Supplementary Fig. S2). A previously unidentified pocket, from here onward referred to as DIM pocket, was identified at the interface region of the AR LBD homodimer. It is located near the conserved contacts of the hinging motions and could be used to design dimerization inhibitors that disrupt AR LBD-LBD dimerization (Fig. 1).

The interactions within the DIM pocket were probed using FTmap leading to a receptor-based pharmacophore model. Using MOE the pharmacophore model was used to query a 3D-conformational databases of commercial available drug-like molecules in search for molecules fitting the shape and interaction within the DIM pocket (Supplementary Fig. S2). Over 6,000 pharmacophore-based hits that matched the pharmacophore query were docked into the DIM pocket. The docking poses were then subjected to a pharmacophore-based postfilter which only retains those docked poses agreeing with the intended interactions from the initial pharmacophore query. In the visual evaluation stage, 249 compounds were evaluated for additional complementarity not described within the pharmacophore query and chemical drug-likeness (PAIN; refs. 31, 32). Finally, 29 potential compounds (DIM1 to DIM29) were selected and acquired on the basis of the visualization of the DSX per-atom contributions (24) and good interactions with binding site residues.

The antagonistic activity of the 29 DIM compounds (1 and 10 μmol/L) on the AR was tested in a previously described stable luciferase reporter HEK 293 (ClARE) cell line in presence of 1 nmol/L DHT (Supplementary Table S1; ref. 25). Three compounds (DIM20, DIM24, and DIM25) with similar structures displayed inhibition of DHT-induced AR activity with limited cytotoxic effects (Fig. 1C; Supplementary Table S1). The predicted binding modes of these compounds matched the pharmacophore query features in the DIM pocket, leading to a clash with the second monomer in the AR LDB dimer (Supplementary Fig. S2C–S2E).

The chemotype of DIM20 was chosen for a structure–activity relationship (SAR) by catalog search. DIM20 is superior over DIM24 and DIM25 because of its lack of cytotoxic effects and its activity at both concentrations (10 and 1 μmol/L). DLS confirmed solubility of the compounds at the evaluated concentrations. The DHT-induced AR activity in the ClARE cells is reduced to respectively 18% (10 μmol/L) and 60% (1 μmol/L) relative to 1 nmol/L DHT (100%; Supplementary Table S1). Derivatives of DIM20 were identified from commercial vendors using the MOE Molecular Fingerprint Similarity Search function with a 4-point pharmacophore based fingerprint (FP:piDAPH4). The search resulted in 55 additional derivatives belonging to the same chemotype but containing different scaffolds or decorations on the 2- and 3- positions (Supplementary Fig. S3).

Ten DIM20 analogs inhibited the DHT signal in the ClARE cells to less than 60% at 10 μmol/L and were further used to determine IC₅₀ and CC₅₀ (Supplementary Fig. S4; Supplementary Table S1). DIM20 however remained the most active and least toxic compound within its chemotype suggesting that the virtual screening procedure indeed identified the most promising derivative from the commercial chemical space.

Specificity of the DIM20 family for AR over the other steroid receptors, was tested in a universal steroid receptor luciferase reporter assay both in agonistic and antagonistic settings (Supplementary Figs. S5 and S6; ref. 33). No agonistic activity of the DIM compounds was detected on any of the receptors tested (Supplementary Fig. S5). DIM20 and DIM20.39 preferentially inhibit the AR with minor antagonistic effects on estrogen receptor alpha and/or beta and moderate effects on the progesterone receptor (PR; Supplementary Fig. S6). A dose-dependent decrease of PR activity was observed for DIM20, DIM20.39 but also for Enz (Supplementary Fig. S6). This was examined further by comparing the AR LBD with the PR LBD structures computationally. The DIM pocket was also identified on the PR and it shows high conservation among AR and PR, except for A748/G763 and G683/D697. The latter two residues however point away from the binding site leaving the direct contacts with the LBD protein backbone identical (Supplementary Fig. S7A). The binding modes of DIM20 and DIM20.39 in the DIM pocket of PR were predicted via docking simulations. In general, they occupied both DIM pockets (AR vs. PR) with the same oriented modes (Supplementary Fig. S7B). This prediction could explain the antagonistic properties of the DIM compounds on PR. Moreover, the molecular docking of Enz in the PR LBD in antagonistic conformation (PDB ID: 2ovh) indicated that it can bind the LBP of the PR in the same manner as Asoprisnil, a selective progesterone receptor modulator (Supplementary Fig. S7C; ref. 34). Both ligands made hydrogen bonds with Q725 and C891. This sequence similarity between AR and PR can explain why not only the DIM compounds but also Enz can behave as an antagonist to both receptors.

For proof of concept, we tested the ability of the compounds to inhibit DHT-induced dimerization of both the AR fl (Fig. 2A) and of the isolated LBD (Fig. 2B). Using the NanoBiT complementation assay, we demonstrate AR fl dimerization and LBD dimerization at 10 nmol/L of DHT. Treatment with DIM compounds reduced the amount of dimerization that occurred upon addition of DHT. The orthosteric antagonist Enz was used as positive control as it was shown earlier to inhibit LBD dimerization in abFRET (5). To exclude that the DIM compounds have an effect on the binding of DHT in the settings of the dimerization assay, we performed a whole-cell competition assay with ³H-labeled DHT (Fig. 2C). The presence of 5 μmol/L of DIM20 or DIM20.39 did not affect the binding of ³H-DHT to the AR (P > 0.05), while 25 μmol/L of DIM20 or DIM20.39 increased the EC₅₀ for ³H-DHT by only 2-fold (P = 0.024) or 3.6-fold (P < 0.001), respectively (Fig. 2C). A total of 10 μmol/L Enz, on the other hand, introduced a 64-fold increase in the EC₅₀ for ³H-DHT (P < 0.001) underlining its competitive binding nature (Fig. 2C).

BLI was applied to confirm binding of the DIM compounds to the AR LBD in presence of high levels of DHT (500 nmol/L). For attachment to the SSA sensor, an avi-tag was integrated at the carboxyterminal end of the LBD. By bacterial coexpression with BirA, this avi-tag is in situ biotinylated. Reversible binding to the AR LBD was demonstrated for both DIM20 and DIM20.39 with apparent K_Ds of 35.7 μmol/L and 29.8 μmol/L, respectively (Fig. 3). The AR20-30 peptide that binds the AF-2 surface of the LBD was used as positive BLI control and demonstrated an apparent K_D of 0.94 μmol/L.

As a first step in the translation of DIM compounds to AR therapeutics for prostate cancer, we determined their antiproliferative effect on several AR-positive prostate cancer cell lines. The VCaP cell line, which is derived from a vertebral metastatic prostate cancer lesion, expresses high levels of AR due to the amplification of the AR gene (35). The DIM compounds significantly reduced VCaP cell proliferation when compared with the vehicle condition (Fig. 4A). This effect, which was concentration dependent, was most evident after 6 days.

When following up the proliferation of LNCaP cells that express the AR T877A variant (36, 37), the DIM compounds inhibit proliferation as monotherapy (Fig. 4B). Combination therapy of DIM20 and Enz leads to synergistic effects in five of six of the drug combinations while all six drug combinations of DIM20.39 and Enz demonstrated synergistic effects (Fig. 4B; Supplementary Table S2). As opposed to the AR-positive prostate cancer cell lines, DIM20 and DIM20.39 only marginally affected the proliferation rate of AR-negative cell lines such as HEK 293 and PC3 (P > 0.05; Fig. 4C).

Proliferation of the 22rv1 cell line, that expresses both the AR H874Y variant and a truncated ARV (38–40), is not affected by Enz (P > 0.05; Fig. 5A). A dose-dependent inhibition of 22rv1 proliferation is however observed for DIM20-treated cells albeit at a nonsignificant level (P > 0.05; Fig. 5A). Of note, transcription of a luciferase reporter regulated by heterodimers of AR fl and truncated AR is significantly reduced by DIM20 as well (Fig 5B).

To investigate their molecular mechanism of action, we tested DIM compounds in combination with other AR-targeting compounds. We studied the effect of antagonists in combination with a gradient of DHT on the DHT-induced AR activity of the ClARE cell line. A constant amount of DIM compound (10 μmol/L DIM20 or 25 μmol/L of DIM20.39) resulted in a reduction of the maximal activity induced by DHT, while a constant amount of Enz (1 μmol/L) retained the maximal activity but severely increased the EC₅₀ of DHT (Fig. 6A). For TRIAC (10 μmol/L), a noncompetitive antagonist for AR that binds in the BF3 pocket, we observed a similar reduction of the maximal AR activity induced by DHT as was observed for the DIM compounds (Fig. 6B). Combining DIM20 with TRIAC leads to an even further reduction of the maximal AR activity indicating a different binding site and mechanism of action.

We further investigated whether the combination of the DIM compounds with Enz could act synergistically on the AR by cotreatment of ClARE cells with doses corresponding to 0.25×IC₅₀, 0.5×IC_5O, IC₅₀, 2×IC₅₀, 4×IC₅₀ with IC₅₀ of DIM20 and Enz previously determined to be 1.30 μmol/L and 0.125 μmol/L, respectively (Fig. 6C). The combination index plot for this cotreatment experiment demonstrates synergism for DIM20 and Enz with the level of synergism increasing with the treatment dose (Fig. 6C).

The AR is a well-studied transcription factor because of its prominent role in prostate cancer progression and treatment. Inhibition of the AR via androgen deprivation therapy or AR antagonists is the cornerstone of prostate cancer therapy (4). AR alterations are well-known resistance mechanisms that limit the success of current treatment options (6). During resistance, the dependence of the tumor on AR remains, which opens up the possibility for new, alternative AR-targeted treatments. Here we provide proof of concept that LBD dimerization can be inhibited with small molecules and that this is an effective mechanism to inhibit AR-dependent proliferation of prostate cancer cells.

The biological relevance of the LBD dimerization surface that was apparent in the AR LBD crystal structure (PDB ID: 5jjm), was confirmed by MD simulations and several key interactions (W751 and T755) were identified. Disruption of LBD dimerization by the W731R mutation (corresponding to W751R in human AR) clearly affects activity of the AR as evidenced by the androgen insensitivity of the AR^Lmon mouse model (12). After defining targetable sites within the LBD dimerization interface, a pocket surrounded by H1, L1-L3, and H5 was defined as the DIM pocket. The DIM pocket is dynamically stable with a low Root-Mean-Square-Fluctuations–based B-factor value (Supplementary Fig. S2B) providing a useful platform for receptor-based drug design.

An in silico screening identified potential LBD dimerization inhibitors that would fit the DIM pocket from a large chemical library of compounds. Three compounds (DIM20, DIM24, and DIM25) that share structural elements were found to inhibit AR activity. This DIM family was further extended with 55 derivatives of which 10 compounds demonstrate AR inhibition in the μmol/L range. A further SAR by catalog search did not yield derivatives with a better inhibitory profile compared to DIM20, meaning that the virtual screen already brought us the most interesting hits. Further efforts to optimize the affinity of the DIM compounds via rational synthesis will result in a promising, new and alternative mode of AR inhibition. Soaking experiments of the DIM compounds in AR LBD-DHT crystals using previously reported conditions (41) failed, potentially due to limited accessibility of the DIM site within the crystal packing.

The chemotype of the DIM compounds consists of a bicyclic core (imidazo[1,2-a]pyrimidine in case of DIM20) decorated at its 2- and 3- position (with methoxyphenol and methoxyaniline, respectively, in case of DIM20; Supplementary Fig. S3). The SAR by catalog further expanded the DIM20 family but also pointed out that some modifications of the DIM20 chemotype are not allowed. In general, there is no clear trend linking a functional group with an active or inactive derivate. However, introducing hydrophobic functional groups including methyl, ethyl, isopropyl, and methoxide on the bicyclic core or the 3- position seems to impair the antagonistic properties (Supplementary Fig. S3B). This can be explained by the spatial effect or sterical hindrance the derivative introduces. However, this explanation is not fully applicable to all cases. All modifications that increased the size or removal of the 3- position are associated with inactive chemotypes (Supplementary Fig. S3C). These derivatives are either too big to occupy the DIM pocket or they lose the aromatic function which is important for activity. This reasoning seems to hold true for modifications introduced in the bicyclic core (Supplementary Fig. S3D). Most of the derivatives’ cores are bigger than that of DIM20, thus a spatial effect plays an important role in this change. This correlates with the fact that according to our virtual screening the ligands are buried deep within the DIM pocket.

In-depth analysis of the AR inhibitory activity of DIM20 and DIM20.39 was carried out with a focus on the AR dimerization potential, AR-dependent proliferation of prostate cancer cell lines and the interplay between mechanistically distinct AR inhibitors.

Several experiments support the DIM20 family of compounds to be noncompetitive AR-LBD targeting inhibitors. First of all, reversible and concentration-dependent binding of DIM20 and DIM20.39 to the AR LBD was demonstrated in presence of high levels of DHT, indicating a different binding site for the DIM compounds compared with DHT (Fig. 3).

Second, DIM20 and DIM20.39 reduce intracellular dimerization of the AR fl as well as dimerization of the isolated LBDs (Fig. 2A and B), while exhibiting only limited effect, if any, on ³H-DHT binding at these conditions (Fig. 2C).

Third, both DIM20 and DIM20.39 show the typical profile of a noncompetitive inhibitor based on their effect on a DHT dose–response curve for AR activity in ClARE cells (Fig. 6A). Noncompetitive inhibitors typically lower the maximal response of the agonist curve by occupying a separate binding site on the receptor, thereby inhibiting its activity (42). A competitive antagonist such as Enz, on the other hand, leads to a shift of the EC₅₀ for DHT toward higher concentrations (Fig. 6A; ref. 42). DIM20.39 reduces both the maximal activity and induces a shift of the EC₅₀ to higher concentrations which indicates that DIM20.39 can also bind to the LBP at higher concentrations, or that it allosterically influences the DHT-binding capacity in the LBP by binding the DIM pocket (Figs. 2C and 6A). DIM20 showed a similar noncompetitive profile as TRIAC, which binds the BF3 region of AR (Fig. 6B).

Finally, we observed that cotreatment with DIM20 and Enz leads to synergism as observed on the AR transcriptional output in the ClARE cell line (Fig. 6C), but also on the proliferation of androgen-responsive LNCaP cells (Fig. 4B).

In conclusion, the DIM compounds act as noncompetitive antagonists on the AR LBD which favors their use over the current clinically used orthosteric AR antagonists (Enz, apalutamide, and darolutamide) in case of resistance due to acquired mutations in the AR LBP. Moreover, a different site of binding also enables cotreatment with these current AR inhibitors to ensure a prolonged or more extensive AR inhibition.

The expression of truncated ARVs that lack the LBD has been proposed as potential resistance mechanism against the classical AR antagonists. However, in clinical samples, the ARVs are always coexpressed with AR fl, with the latter being the dominantly expressed variant. Under these conditions, ARVs are known to acts as heterodimers with AR fl (9, 43). Importantly, these heterodimers depend on AF-2. More precisely, the specific interaction between the AF2 of the AR fl and the ²³FQNLF²⁷ motif of the ARV is deemed essential for heterodimer formation (44, 45). Importantly, the classical orthosteric AR antagonists prevent protein–protein interactions between the AF2 site and coregulators (46). Whether this is also true for inhibitors of the LBD DIM pocket remains to be investigated. On the basis of the inhibitory effect of the DIM compounds on the transcriptional output of AR fl–ARV heterodimers (Fig. 5B) and the inhibition of the proliferation of the 22rv1 cells (Fig. 5A), this study indicates that DIM20 could even be effective in prostate cancer cells which coexpress AR fl and truncated ARVs.

Although the DIM compounds were designed for the AR, they are also inhibiting the PR. The almost complete conservation of the AR DIM pocket in the PR suggests a conserved role for LBD dimerization for PR and AR. This cross specificity could however be a general phenomenon for AR inhibitors as we also observed it for Enz that could fit in the LBP of PR. Inhibitory effects of the DIM compounds on PR are not expected to be harmful to male patients with prostate cancer, because PR mainly has a function in the female physiology (47). PR inhibition might even be beneficial because the PR-B has been linked with disease progression in prostate cancer, although further research is required (48).

This study supports the possibility of inhibiting AR function via the disruption of LBD dimerization by small molecules. First of all, further structural optimization of the DIM20 chemotype is required to improve the potency for AR inhibition. Because of its noncompetitive binding, the DIM20 family could complement the current treatments with AR inhibitors that bind the LBP. We also show that the DIM compounds remain active when ARVs are expressed in tumors resistant to currently approved AR signaling inhibitors. The discovery and further characterization of the DIM20 family suggest that disruption of LBD dimerization is a good therapeutic strategy for AR-related disease. Consequently, it may be useful to explore LBD dimerization as a target for other nuclear receptor-associated diseases.

C. Helsen reports a patent for EP22196697.1 pending. T.T. Nguyen reports a patent for EP22196697.1 pending. J. Schymkowitz reports personal fees and other support from Aelin Therapeutics outside the submitted work. F. Claessens reports a patent for EP22196697.1 pending. A. Voet reports a patent for EP22196697.1 pending. No disclosures were reported by the other authors.

C. Helsen: Formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. T.T. Nguyen: Software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. X.Y. Lee: Formal analysis, methodology, writing–original draft, writing–review and editing. R. Eerlings: Formal analysis, methodology, writing–original draft, writing–review and editing. N. Louros: Investigation, writing–review and editing. J. Schymkowitz: Funding acquisition, writing–review and editing. F. Rousseau: Funding acquisition, writing–review and editing. F. Claessens: Conceptualization, supervision, funding acquisition, writing–original draft, writing–review and editing. A. Voet: Conceptualization, software, supervision, funding acquisition, writing–original draft, writing–review and editing.

The authors thank Sofie De Block and Hilde De Bruyn for their excellent technical assistance. We are thankful to Carlotta Borgarelli, Gert Steurs, and Prof. Wim De Borggraeve for their help with NMR. Financial support for this study was obtained from KU Leuven Internal funding (C14/17/067 granted to A. Voet and C. Helsen) and from the Research Foundation Flanders (FWO; G094018N granted to F. Claessens and to C. Helsen). The Switch Laboratory was supported by the Flanders Institute for Biotechnology (VIB; C0401 to F. Rousseau and J. Schymkowitz) and the Research Foundation Flanders (FWO; infrastructure grant G0H1716N to J. Schymkowitz) and Postdoctoral Fellowships 12P0919N and 12P0922N to N. Louros).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.