SEARCH

SEARCH BY CITATION

Keywords:

  • 11β-Hydroxysteroid dehydrogenase;
  • Enzyme inhibitor;
  • Model refinement;
  • Pharmacophore model;
  • Virtual screening

Abstract

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Methods
  5. 3 Results
  6. 4 Discussion
  7. 5 Conclusions
  8. Acknowledgements
  9. Supporting Information

11β-Hydroxysteroid dehydrogenases (11β-HSD) control the intracellular concentrations of glucocorticoids: 11β-HSD1 converts the inactive cortisone to the active cortisol, and 11β-HSD2 is responsible for the opposite reaction. Inhibition of 11β-HSD1 is beneficial in the treatment of metabolic syndrome, whereas 11β-HSD2 inhibition leads to hypertension. Therefore, 11β-HSD1 inhibitors should be selective over 11β-HSD2. To support drug discovery and toxicological studies, we have previously reported pharmacophore models for 11β-HSD1 and 2 inhibition. These models represent the common chemical features of 11β-HSD inhibitors, which were used as virtual screening filter. Since new inhibitors are constantly discovered, the quality of the pharmacophore models has to be evaluated in order to maintain a good predictive power. In this study, we report a systematic evaluation and refinement of our pharmacophore model collection. We employed our models for virtual screening, especially focusing on the 11β-HSD2 inhibition. In total, 42 compounds were biologically evaluated and among these we discovered 17 11β-HSD inhibitors that decreased the residual enzyme activity to 50% or less at the concentration of 20 µM. The experimental 11β-HSD1 and 2 readouts from these compounds were used for further model refinement. Evaluation metrics were applied for a quantitative comparison of the old and newly generated models which resulted in a set of improved pharmacophore models offering reliable in silico tools for the identification of novel and selective 11β-HSD inhibitors.


1 Introduction

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Methods
  5. 3 Results
  6. 4 Discussion
  7. 5 Conclusions
  8. Acknowledgements
  9. Supporting Information

11β-Hydroxysteroid dehydrogenases (11β-HSDs) control the intracellular availability of active glucocorticoids.[1] 11β-HSD1 is a NADPH-dependent reductase, converting inactive cortisone to its active 11-hydroxy form cortisol (Figure 1). The NAD+-dependent 11β-HSD2 is responsible for the inactivation of cortisol back to its keto form cortisone. 11β-HSDs belong to the superfamily of short chain dehydrogenases/reductases (SDR), which is a large enzyme family with a broad spectrum of substrates and catalyzed reactions.2 Although the enzymes in this family have low sequence identities, their 3D-structures show high similarity, making them well superimposable. Like all the enzymes in this superfamily, 11β-HSDs share a so-called Rossman fold, where six parallel β-sheets are surrounded by four α-helices.2b

thumbnail image

Figure 1. Interconversion of cortisone to cortisol by 11β-HSDs.

Download figure to PowerPoint

11β-HSD1 is expressed especially in the liver, ovaries, adipose-, and adrenal tissues, all of them being target tissues of glucocorticoids.3 In contrast, 11β-HSD2 is mainly localized in mineralocorticoid-targeted tissues such as kidneys and colon, where it protects the mineralocorticoid receptors from cortisol excess.4 Cortisol binds to the mineralocorticoid receptors with comparable affinity than the main ligand aldosterone, but its circulating concentration is 100–1000 times higher than that of aldosterone.5 Normally, 11β-HSD2 rapidly inactivates cortisol to cortisone, keeping the mineralocorticoid receptors unaffected. Impaired 11β-HSD2 activity leads to excessive cortisol-dependent mineralocorticoid activation and severe hypertension.

Overexpression of 11β-HSD1, especially in adipose tissue, can cause the symptoms of metabolic disorder: visceral obesity, hyperglycemia, insulin resistance, and increased serum fatty acid and triglycerides levels.6 High concentrations of cortisol only in the liver does not cause obesity or central adiposity, but steatosis, hypertension, mild insulin resistance, and modest dyslipidemia.7 Stress or obesity does not lead to hyperglycemia if 11β-HSD1 is lacking.8 Furthermore, overexpression of 11β-HSD1 and therefore high concentrations of active cortisol in adipose tissue can cause full metabolic disorder. Thus, inhibition of 11β-HSD1 is a promising way to treat this condition. It is important to avoid simultaneous inhibition of renal and vascular 11β-HSD2, because lack or inhibition of this enzyme will lead to hypokalemia, hypertension, edema formation and renal enlargement.9 Inhibition of 11β-HSD2 may be responsible for some of the adverse cardiovascular effects of anabolic steroid10 and natural compounds such as the licorice constituent glycyrrhetinic acid.11 However, selective 11β-HSD2 inhibitors provide a pharmacological tool and topical applications may offer a potential treatment against hyperkalemia for hemodialysis-patients.12

A promising approach to the challenging task of inhibitor development is pharmacophore modeling. Pharmacophore models represent the electrostatic and steric features that are necessary for an optimal interaction of a small molecule with its biological target (enzyme, receptor).13 A pharmacophore model shows the optimal 3D-arrangement of hydrogen bond acceptors (HBA) and donors (HBD), hydrophobic areas (H), aromatic rings (Ar), ionizable groups (PI/NI), and metal binding fragments for triggering or blocking a specific biological effect. Exclusion volumes (XVOL) – forbidden areas – or a shape may be added to represent the size and shape of the ligand binding pocket. Pharmacophore modeling and pharmacophore-based virtual screening are powerful tools to achieve an enrichment of active compounds from a chemical database in comparison to a random screening.14

We have previously published pharmacophore modeling studies on 11β-HSD enzymes.15 In the first study two ligand-based models were developed, one for selective 11β-HSD1 inhibition (Model 1) and one for non-selective inhibition, preferring 11β-HSD2 (Model 2) (Figures 2 A and 2 B). During the development of these models their performance was tested by screening a stockroom database which consisted of 144 compounds. Model 1 predicted 17 compounds, of which 12 showed a moderate or potent inhibition of 11β-HSD1. In contrast, Model 2 identified two hits, both potent non-selective 11β-HSD inhibitors. Both models were then experimentally validated by screening several commercially available databases; hits were analyzed and selected hit molecules were tested against 11β-HSD activity. Seven from the 30 tested compounds inhibited 11β-HSD1 more than 70% at the concentration of 10 µM. Two of these compounds were picked by Model 1 and five by Model 2. When counted together, this represents 23% of success rate.

thumbnail image

Figure 2. Pharmacophore models for 11β-HSD inhibitors: The original 11β-HSD1 model (Model 1, A) consists of four H features, one HBA, one HBD, and a shape query. The non-selective 11β-HSD inhibitor model (Model 2, B) is composed of five H features, four HBAs, and a shape. The 11β-HSD2 model (Model 3, C) consists of three HBAs, two H features, one HBD, and XVOLs. The pharmacophoric features are labeled as following: hydrophobic – H, hydrogen bond acceptor – HBA, hydrogen bond – HBD. The shape queries and XVOLs are not labeled. Two different pharmacophore generation programs were used for the development of the pharmacophore models: Catalyst (Models 1 and 2) and LigandScout (Model 3)

Download figure to PowerPoint

In the course of our study on triterpenoids inhibiting 11β-HSD2, we docked some 11β-HSD2-selective inhibitors to the active site of our 11β-HSD2 homology model.15b Based on these docking solutions and the homology model, we derived a pharmacophore model for selective 11β-HSD2 inhibition (Model 3). Because Model 3 was derived only from triterpenoid inhibitors, its use was limited and did not find other scaffolds.

Rollinger et al.16 reported constituents from Eriobotrya japonica that inhibit 11β-HSD1. In this study, a flipped binding mode, where certain triterpenoids would not form interactions with catalytic amino acids but with Thr124 and Tyr177 was suggested. In addition, a pharmacophore model for triterpenoids inhibiting 11β-HSD1 was suggested. The model consists of three HBAs placed on the hydroxyl-groups on the triterpenoid’s positions 2, 3, and 28, and nine Hs placed on the triterpenoid core. In fact, this pharmacophore model is comparable to Model 3: Whereas the 11β-HSD1 selective triterpenoids have only HBA-functionalities, the 11β-HSD2 selective ones bear as well an HBD-group.

In order to maintain a good predictive pharmacophore model quality and performance, the models have to be re-evaluated regularly. In this study, we report an evaluation and the refinement of our 11β-HSD pharmacophore models applying the workflow depicted in Fig. 3. To the best of our knowledge, this kind of systematic refinement study has not been reported before. To get more data for this model refinement work, we employed all our 11β-HSD models to a prospective screening of several databases. Enzyme activity tests were performed for 42 compounds and data obtained from these tests were used in an iterative scale for model refinement. Finally, evaluation metrics were applied to the old and new models for a quantitative comparison of their performances.

thumbnail image

Figure 3. Schematic overview of the workflow.

Download figure to PowerPoint

2 Methods

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Methods
  5. 3 Results
  6. 4 Discussion
  7. 5 Conclusions
  8. Acknowledgements
  9. Supporting Information

2.1 Database Generation and Virtual Screening

2.1.1 11β-HSD Inhibitors Literature Database

A database of 11β-HSD inhibitors from literature was assembled (Supporting Information Table S-1). These compounds were transformed into a 3D-database using Discovery Studio 3.0. For each compound, maximum 255 conformers were calculated using BEST algorithm.

2.1.2 Innhouse Database and Commercial Databases

The database composed of compounds stored at the University of Innsbruck, the so-called Innhouse database, consists of 2706 compounds. This database includes drugs, reagents, and natural compounds that are on stock at the Institute of Pharmacy, University of Innsbruck, Austria. The conformations of the 3D-structures were generated using Discovery Studio 3.0 (Accelrys Software Inc. 2009–2012). For each compound, a maximum of 100 conformations was set during the conformational analysis using the FAST algorithm. In DiscoveryStudio, these conformations were transformed into a 3D database using the “build 3D-database” protocol. Within LigandScout (www.inteligand.com,17), the Innhouse database was calculated with the idbgen tool of LigandScout using Omega-fast settings (25 conformations/molecule). The previously reported natural products database DIOS (n=9676)18 and the endocrine disrupting chemicals (EDC) database (n=76677)19 were prepared as described previously. The commercial databases Specs (n=199564) and Maybridge (n=60542) were transformed into LigandScout databases using the idbgen-tool and Omega-fast settings.

2.1.3 Prospective Virtual Screening

For screening the DIOS database with Model 1 as well as the screening of the EDC library with Model 2, Catalyst 4.11 (www.accelrys.com) was employed. For all further screenings, Discovery Studio 3.0 was used. The Innhouse database was screened with Models 1, 2, and 4 using the “search 3D-database” protocol with BEST settings. Screening of the Innhouse, Specs, and Maybridge databases with the Models 3 and 5 were performed using LigandScout3.0a. One omitted feature was allowed during the screening experiments with Models 3 and 5. The use of different modeling and screening programs is advisable because this often results in complementary models which all find active compounds without much hitlist overlap.20 Since the models were originally constructed and validated using two distinct programs that use different algorithms for conformation generations, pharmacophore model construction, and virtual screening,17,21 the further screening and validation had to be continued using the two different software packages.

2.1.4 Refinement Database with Newly Tested Compounds

The refinement of Models 1, 2, and 4 was performed with Discovery Studio 3.0. For this purpose, a 3D-database of the biologically tested compounds was composed with the “build 3D-database” protocol of Discovery Studio. Using BEST settings, 250 conformations for each molecule were calculated. For all the other options, the default settings were kept. To test the model performance in each refinement step, this database was screened using the “search 3D-database” tool of Discovery Studio. The search method was set to BEST and the minimum interfeature distance to 0, which means that two neighboring features can be mapped directly next to each other. Otherwise, the default settings were used.

Models 3 and 5 were refined using LigandScout3.03a. For the refinement of these models, a test set database was composed using the idbgen-tool of LigandScout with Omega-best (500 conformations/molecule) settings.

2.1.5 Validation of the Refined Models with Literature Data

For further independent validation of the refined models, a database of compounds, which activities against 11β-HSDs was known, was composed (Supporting Information, Table S-3). This database consisted of 62 compounds, from which 31 were 11β-HSD1-selective inhibitors, six 11β-HSD2[BOND]selective inhibitors, five nonselective ones, 11 inactive compounds against both 11β-HSDs, and 9 entries, which were inactive against 11β-HSD1, but which activity against 11β-HSD2 was not published. These compounds were selected from the literature and generated using ChemBioDraw Ultra 12.0 (1986–2010 CambridgeSoft).

All the compounds were transformed into 3D multiconformational databases using the following settings: In DiscoveryStudio, the database was constructed using the Build 3D-database[BOND]protocol. During the database generation, a maximum of 250 conformations was calculated for each molecule using the BEST algorithm. Within LigandScout, the database was generated using the idbgen-tool and Omega-best settings (a maximum of 500 conformations/molecule). The virtual screenings of this database were performed as described in Section 2.1.3.

2.2 Biological Assays

Biological testing of virtual hits was carried out as described earlier.15b Briefly, lysates of HEK-293 cells stably expressing human recombinant 11β-HSD1 were incubated for 10 min at 37° in a total volume of 22 µL of TS2 buffer (100 mM NaCl, 1 mM EGTA, 1 mM EDTA, 1 mM MgCl2, 250 mM sucrose, 20 mM Tris[BOND]HCl, pH 7.4) containing 200 nM radiolabeled cortisone, 0.5 mM NADPH and either vehicle or the corresponding inhibitor. Stock solutions of all inhibitors were prepared in methanol or dimethylsulfoxide to a final concentration of 20 mM. Inhibitors were diluted in TS2 buffer and immediately used. Reactions were terminated by addition of unlabeled cortisone and cortisol in methanol. Steroids were separated by TLC, followed by analysis of conversion of cortisone to cortisol by scintillation counting. Vehicle served as negative control, glycyrrhetinic acid as positive control.

11β-HSD2 dependent oxidation of cortisol to cortisone was measured similarly for 10 min at 37° using lysates of HEK-293 cells stably expressing human 11β-HSD2 and using radiolabeled cortisol at a final concentration of 50 nM and 0.5 mM NAD+.

2.3 Pharmacophore Model Refinement

The model refinement in all of the cases was first approached by analyzing the model quality: which of the models needed improvement in specificity, and which ones in sensitivity. In general, an improved specificity was aimed by adding additional spatial restrictions in forms of shapes and/or XVOLs. In general, the shape itself is formed from the 3D-coordinates of the query molecules atoms, each of them with a corresponding radius. During the screening, a shape for each hit molecule candidate is generated and aligned with the query, their intersection and union are approximated. The volume of the intersection divided by the united volume of the aligned shapes gives the similarity of the shape. Sensitivity was increased by (i) removing all spatial restrictions from the model, (ii) screening the testset and evaluating which active compounds could not map which pharmacophore feature, and (iii) deleting features and set some optional, that were not mapped by many active testset compounds and/or adjust the size of features so that active compounds could still map them.

The model performance before and after the refinement was analyzed by counting the number of fitting actives (true positives), fitting inactives (false positives), discarded actives (false negatives) and discarded inactives (true negatives) for each model. These numbers were used for calculating the descriptive parameters: yield of actives (YA) (Equation 1), enrichment factor (EF) (Equation 2), sensitivity (Equation 3), and specificity (Equation 4).(1), (2), (3), (4)

  • equation image(1)
  • equation image(2)
  • equation image(3)
  • equation image(4)

3 Results

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Methods
  5. 3 Results
  6. 4 Discussion
  7. 5 Conclusions
  8. Acknowledgements
  9. Supporting Information

3.1 Model Refinement Using Literature Data (Model 4 and Model 5)

After the publication of the first 11β-HSD models, several new inhibitors from various chemical scaffolds have been reported (Supporting Information Table S-1). Therefore, Model 1 was re-evaluated for its correct retrieval of novel inhibitors. Since most of the selective 11β-HSD1 inhibitors do not have a HBD (e.g.,22 and Supporting Information Table 1), the HBD feature of Model 1 was exchanged with an HBA. Inspections of the binding modes of the cocrystallized inhibitors of 11β-HSD1 revealed that some of the compounds, e.g. the arylsulfonylpiperazine inhibitor (PDB-code 3czr23), have a V-shaped binding mode. The initial shape of Model 1 was not able to accommodate these compounds; accordingly the shape constriction was removed. This updated 11β-HSD1 model (Model 4) (Figure 4 A) was able to find 13 of the 14 11β-HSD1 inhibitors from the database that was used for the early refinement of this model (Supporting Information, Table S-1). In contrast, only two of the compounds from this database fitted to the Model 1.

thumbnail image

Figure 4. Pharmacophore Models 4 (A) and 5 (B), refined based on literature data. Model 4 consists of four H and two HBA features. The additional 11β-HSD2 model, Model 5, consists of three H, two HBA, and one HBD features as well as XVOLs. The pharmacophoric features are labeled as following: hydrophobic – H, hydrogen bond acceptor – HBA, hydrogen bond – HBD. The XVOLs are not labeled. Model 4 was refined using Discovery Studio and Model 5 using LigandScout.

Download figure to PowerPoint

Due to the lack of known selective 11β-HSD2 inhibitors, the refinement of Model 2 was postponed for later. Since the previously published 11β-HSD2 model (Model 3) showed restrictivity towards diverse chemical scaffolds and strongly favored triterpenoids,15b a second 11β-HSD2 model was developed (Model 5) (Figure 4 B). Thus, the model was only based on triterpenoids and the homology model of 11β-HSD2.15b All the reported active triterpenoids bear the same keto-group in 11-position of the triterpenoid core; however, that feature may not be necessary for the selectivity towards 11β-HSD2.16 Therefore, the central HBA feature indicating the position of the 11-keto group (Figure 1 C) was removed. Further analyses of common features of active compounds led to the addition of a H feature. The HBA and the HBD features were moved 1 Å closer to each other, which better represented the actual distances of the corresponding chemical groups present in active structures. In addition, the positions of H features were adjusted.

3.2 Prospective Virtual Screening and Selection of Hits

Models 1–5 were used to screen the Innhouse database, DIOS database, EDC library, Specs, and Maybridge databases. The derived numbers of hits are given in Table 1.

Table 1. The number of virtual screening hits in all used databases. The number is represented in brackets if one omitted feature was allowed during the screening. ns: not screened.
DatabaseModel 1Model 2Model 4Model 3Model 5
Innhouse3472570 (2)4 (49)
DIOS172nsnsnsns
EDCns31nsnsns
Specsnsnsns1349
Maybridgensnsns(115)(212)

Hits for biological evaluation were selected based on their geometrical pharmacophore fitness score, novelty, and availability. Seventeen compounds were selected from the Innhouse database. Because there are only few known 11β-HSD2-selective inhibitors,15b, 24 most of them triterpenoids, we decided to emphasize on novel scaffolds by focusing on structurally diverse hits from commercial databases. Fourteen synthetic compounds were selected for biological testing: ten from the Maybridge database and four from the Specs database, respectively. From the EDC library, four compounds were selected for biological evaluation. To additionally investigate promising natural products, two compounds from the DIOS database were biologically tested. Because of compound availability reasons, hispanolone was selected for testing, although it was not directly found by screening but belongs to a frequently found scaffold in the DIOS database. Altogether, 38 compounds were selected for biological studies.

For comparison, additional compounds not identified by a virtual screening were investigated in vitro. Spironolactone was added because of its actions to related targets. Ibuprofen was selected because of its broad usage. Furosemide and gossypol were added as positive controls to the test set because of their already known activities towards 11β-HSD2.24b,25 The test set for the pharmacophore model refinement was composed of these 42 compounds (Table 2 and Supporting Information Figure S-1 for 2D-figures)

3.3 Biological Tests

The 42 compounds were tested against 11β-HSD activity, whereby 7 were 11β-HSD1-selective inhibitors, 7 were selective towards 11β-HSD2, and three turned out as non-selective inhibitors (Figure 5). A compound was considered as inhibitor, if the remaining activity of the respective enzyme was ≤55% at the inhibitor concentration of 20 µM or ≤65% at the inhibitor concentration of 10 µM. The cutoff value was kept relatively high due the lack of known 11β-HSD2 inhibitors. The remaining 25 compounds were considered to be inactive or have very weak activity towards both of the 11β-HSDs.

thumbnail image

Figure 5. 2D-representations of the newly identified 11β-HSD inhibitors with their activities. Activities are announced as percentage of residual enzyme activity at the inhibitor concentration 20 µM, except for octotiamine, lidoflazine, and rifampicin, which were tested at the concentration of 10 µM.

Download figure to PowerPoint

3.4 Pharmacophore Model Refinement

3.4.1 Analysis of Models 1–5

As a first step in model refinement, all the newly tested compounds (Table 2) were fitted into Models 1–5. For each model, the number of true positives, false positives, false negatives, and true negatives were counted. Using these numbers, parameters describing the model quality were calculated. The results are presented in Table 3.

Table 2. Inhibitory activities of the test set compounds.
Compound11β-HSD1[a]11β-HSD2[b]Source DBModel
  1. [a] 11β-HSD1 activity, given as % remaining activity of the enzyme at the inhibitor concentration of 20 µM, unless otherwise stated. [b] 11β-HSD2 activity, given as % remaining activity of the enzyme at the inhibitor concentration of 20 µM, unless otherwise stated. [c] Tested as positive control. [d] Tested because of its broad clinical use.

1 Cefdinir86% (10 µM)103% (10 µM)Innhouse5
2 Ceftriaxone99%104%Innhouse5
3 Cyclosporin97%67%Innhouse4
4 Dobutamine102%97% (10 µM)Innhouse5
5 Ethacrynic acid97%87%Innhouse4
6 Etiroxate90% (10 µM)97% (10 µM)Innhouse5
7 Fenofibrate5%97%Innhouse4
8 Indometacin86%81%Innhouse4
9 Iopronic acid75% (10 µM)99% (10 µM)Innhouse3
10 Ketoconazol67%26%Innhouse4
11 Lidoflazine96%61% (10 µM)Innhouse5
12 Novobiocin101%106% (10 µM)Innhouse5
13 Octotiamine99%52% (10 µM)Innhouse5+3
14 Phtalylsulfathiazol98%108% (10 µM)Innhouse5
15 Rapamycin21%11%Innhouse5
16 Rifampicin96%49% (10 µM)Innhouse5
17 RU2831828%46%Innhouse4
18 AW01077SC82%100%Maybridge5
19 AW011447SC82%100%Maybridge3
20 BTB00686SC87%98%Maybridge5
21 HAN00375SC91%100%Maybridge5
22 HTS01152SC84%100%Maybridge5
23 HTS00335SC84%61%Maybridge3
24 HTS08278SC84%100%Maybridge5
25 KM04146SC22%80%Maybridge5
26 MO00240SC75%61%Maybridge3
27 SRC00931SC87%93%Maybridge5
28 AH478–1440408690%100%Specs5
29 AH487–4148264797%100%Specs5
30 AK968–4183662595%100%Specs5
31 AN645–4338410879%99%Specs5
32 Cephaeline100%74%EDC2
33 Digitoxigenin36%104%EDC2
34 Hecogenin49%82%EDC2
35 Lasalocid84%42%EDC2
36 Hispanolone29%70%DIOS1
37 Marrubiin40%88%DIOS1
38 Monoolein88%52%DIOS1
39 Furosemide18%50%[c] 
40 Gossypol63%23%[c] 
41 Ibuprofen51%95%[d] 
42 Spironolactone78%61%[d] 
Table 3. Test set screening results and calculated descriptive parameters for each model. The numbers in brackets represent the numbers when fitting was performed with one allowed omitted feature.
 11β-HSD1Nonselective model11β-HSD2
Model14235
True positives2/106/100/170(1)/101(5)/10
True negatives29/3211/3225/2532(32)/3229(17)/32
False positives3/3221/320/250(0)/323(15)/32
False negatives8/104/1017/1710(9)/109(5)/10
Yield of actives0.400.2200(1)0.25(0.25)
Enrichment factor1.68/4.20.924/4.20/2.40(4.20)/4.21.05(1.05)/4.2
Sensitivity0.200.6000 (0.10)0.10(0.50)
Specificity0.910.3411(1)0.91(0.53)

Model 1 found two active compounds (hispanolone, marrubiin), but was not able to find the eight other 11β-HSD1 inhibitors (RU283148, rapamycin, digitoxigenin, fenofibrate, furosemide, KM04146SC, hecogenin, and ibuprofen). For some of these active compounds, which do not fit to the model, the fitting would be successful with one omitted feature, either a hydrophobic one or the hydrogen bond donor. As expected in the early re-evaluation of this model (resulting in Model 4), these results also suggest that the hydrogen bond donor feature is not essential for 11β-HSD1 inhibition.

Surprisingly, Model 2 found no hits in the test set screening. A structural comparison of the 11β-HSD inhibitors suggests that the active compounds are unlikely to have so many hydrophobic features or HBAs than Model 2 includes.

As expected, Model 4 performed better than the original Model 1: six active compounds fitted into this model (fenofibrate, KM04146SC, hispanolone, hecogenin, marrubiin, and RU283148). However, 21 inactive compounds also fit to Model 4, in comparison to only three inactives found by Model 1. Exchanging the HBD feature with an HBA and removing the shape improved the sensitivity of the model, but also decreased its specificity. Therefore, the specificity of this model had to be increased with further refinement steps.

Unfortunately there were no compounds that fitted into Model 3. Therefore one omitted feature was allowed for the compound to be a hit molecule. With one omitted feature, one 11β-HSD2 inhibitor, lasalocid, fitted to the model (see also Nashev et al.26). This model showed high specificity and restrictivity.

Model 5 performed better than Model 3. It only recognized one active compound, lasalocid, without any omitted features. Three inactive compounds fitted into the model as well. With one omitted feature, 20 compounds, from which five were identified as 11β-HSD2 inhibitors (rapamycin, lasalocid, gossypol, lidoflazine, and octotiamine) mapped the model.

3.4.2 Model Refinement Steps

Since Model 1 was refined using literature data (Model 4), it was decided to be left out from further refinement steps. Therefore, Model 4 was chosen as a starting point for the refinement of the 11β-HSD1 pharmacophore model. Since it showed already high sensitivity, the only refinement step was to add a shape restriction to increase the specificity of this model. The shape was based on two 11β-HSD1-selective inhibitors: RU28318 and marrubiin. The shape similarity tolerance was set to 0.45, meaning that the compound must show at least 45% similarity in size and shape compared to the combined shape of RU28318 and marrubiin. The extents ratio, which indicates how much the molecule is allowed to differ from the shape in every dimension, was set to be between 0.70 and 1.20. The refined 11β-HSD1 model (Model 4new) consisted of six features (two HBAs and four Hs) and a shape (Figure 6 A).

thumbnail image

Figure 6. The refined pharmacophore models. 11β-HSD1 selective Model 4new (A), non-selective Model 2new (B), 11β-HSD2 Models 3new (C) and 5new(D). The pharmacophoric features are color-coded: HBA – green, H – cyan (A and B) and, HBA – red, H – yellow, XVOL – grey (C and D).

Download figure to PowerPoint

Model 2 was highly restrictive and therefore, in the refinement of this model, the aim was to reduce the restrictivity of the model. A scaffold analysis of the active compounds revealed that most of the actives do not fulfill all of the H features. One of the HBAs may also be left out. In the end, the number of features was decreased by removing one HBA and three H features. For a spatial definition of the maximum hit size, a shape restriction based on rapamycin was added. The shape feature minimum tolerances were set to 0.30, indicating that at least 30% of the rapamycin shape must be filled by a virtual hit. After these steps, Model 2new had five features, three HBAs and two Hs, and a shape (Figure 6 B).

Like Model 2, Model 3 was highly restrictive and not able to identify smaller, structurally diverse 11βHSD2 inhibitors. Therefore, the HBD feature was removed; one of the HBAs was exchanged with an H and set optional. The original XVOLs were removed and a new set of XVOLs based on the positions of the amino acid residues from our homology model15b was added to better represent the shape and size of the binding pocket. After this refinement, Model 3new had five features: two HBAs and three Hs, from which one was set optional (Figure 6 C).

As in the case of Model 3, the specificity of Model 5 had to be improved. This was achieved by deleting an H and the HBD features as well as exchanging one H feature with an HBA. Additionally, the original XVOLs were removed and a set of new ones based on the binding site geometry were added. The refined Model 5new consisted of four features: three HBAs and one H (Figure 6 D). Closer insights in model refinement steps for each model are given in the Supporting Information, Table S-2.

3.4.3 Analysis of Refined Models (2new–5new)

After the refinement steps, the test set database (Table 2) was screened again with the refined models. As previously, the same descriptive parameters were calculated for each model (Table 4).

Table 4. Test set screening statistics of refined models.
 11β-HSD1Nonselective model]11β-HSD2
Model4new2new3new5new
True positives6/105/176/104/10
True negatives21/3218/2526/3225/32
False positives11/327/256/327/32
False negatives4/1012/124/106/10
Yield of actives0.350.380.50.36
Enrichment factor1.47/4.20.939/2.42.10/4.21.51/4.2
Sensitivity0.600.290.600.40
Specificity0.660.720.810.78

Model 4new showed improved specificity, but the sensitivity remained same. The refinement steps of Model 2 increased the sensitivity of the model, whereas the specificity remained high. Both the Models 3new and 5new showed better sensitivity and specificity in comparison to the original Models 3 and 5. As a result, every refined model performed better than the respective unrefined one.

Besides correctly recognizing active compounds, the pharmacophore models for selective 11β-HSD1 and 2 inhibition were able to correctly pick selective compounds. Model 3new found six actives from which four were selective over 11β-HSD1. Similarly, Model 5new correctly recognized four active compounds from which three were selective. Model 4new performed well: out of six active compounds, five were selective over 11β-HSD2. The closer selectivity analysis of the models revealed that even though the Models 3new, 4new, and 5new pick inactive compounds, these inactive compounds are inactive against both 11β-HSDs. This means that these models can predict the 11β-HSD activity correctly.

To further test the model quality on a different data set, a literature-based validation of the models was performed. This literature-based validation database consisted of structurally diverse compounds which have not been used in the development or validation of the models before. The only exception are the 11β-HSD2-selective triterpenoids, which are to the best of our knowledge the only available selective 11β-HSD2 inhibitors. This database was screened with all the unrefined and refined models to see if the refinement of the models was successful. The results of this second validation round were comparable to the results we have got during the model refinement. In all of the cases, the model performance has improved (Table 5).

Table 5. The number of virtual screening hits in the second validation screening.
ModelActivesInactivesSensitivitySpecificitySensitivity/Specificity
10/360/26010
26/425/110.140.550.25
2new12/425/110.290.550.53
30/110/42010
3new7/119/420.640.790.81
415/3616/260.420.381.11
4new11/3613/260.310.50.62
51/112/420.090.950.09
5new6/1110/420.550.760.72

For all but one model, the sensitivity of the models increased, while the specificity decreased or remained the same (Figure 7).

thumbnail image

Figure 7. Differences in the specificities and sensitivities between the original and refined models.

Download figure to PowerPoint

These models showed an improved sensitivity vs. specificity ratio indicating overall model improvement. The sensitivity of Model 4new was lower than the original Model 4. However, also its specificity increased during the refinement, which was beneficial for the overall quality of the model. These results demonstrated that the model quality improvement is often an act of balance between sensitivity and specificity: often a model that is highly sensitive is also lacking specificity and vice versa.

To test the model performance with a bigger database, the Innhouse database was screened with the refined models. All of these models returned with higher number of hits than the unrefined ones: Model 2new with 69 hits, Model 4new with 158 hits, Model 3new with 147 hits and Model 5new with 59 hits. Since the quality of each model is improved, all the models and further virtual screening results serve for reliable starting points for new medicinal chemistry projects.

4 Discussion

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Methods
  5. 3 Results
  6. 4 Discussion
  7. 5 Conclusions
  8. Acknowledgements
  9. Supporting Information

In order to maintain high pharmacophore model quality, a regular model performance analysis is needed, especially in the field where new ligands are constantly developed. As an example, our unrefined models were not able to recognize most of the new 11β-HSD inhibitors. After the refinement, the quality of each model was improved, suggesting better predictability, which was also confirmed by literature-based model validation. However, the lack of structurally diverse 11β-HSD2 inhibitors caused a bias on our data. Most of the known 11β-HSD2 inhibitors are triterpenes. So if a model recognizes this scaffold it usually finds all of the known triterpenoid 11β-HSD2 inhibitors. Therefore, the 11β-HSD2 pharmacophore model quality will have to be tested whenever new inhibitors are available.

In addition to the improved model quality, we have discovered novel active scaffolds especially for 11β-HSD2 inhibition. These new scaffolds aid further research in the field. In addition to these new scaffolds, some already used drugs, such as ibuprofen, ketoconazole and rifampicin are shown to inhibit either of the 11β-HSDs. These findings support the drug development by revealing multitarget activities and serve as a starting point for selective optimization of side activities (SOSA approach).27

Since 11β-HSD1 inhibitors are developed especially against metabolic syndrome and type II diabetes, they have to be selective over 11β-HSD2. Unfortunately, the difference between the binding sites of these two enzymes is unknown. We have previously suggested that selective 11β-HSD2 inhibitors bear a hydrogen bond donor functionality that interacts only with the binding site of 11β-HSD2..[15b] In the light of this pharmacophore model refinement study, this rule can be confirmed only for triterpenoids. However, the refined models correctly picked the right active compounds from the database, and if they found inactive ones, they were inactive against both 11β-HSDs. This suggests that our refined models work as a good tool in predicting selectivities.

5 Conclusions

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Methods
  5. 3 Results
  6. 4 Discussion
  7. 5 Conclusions
  8. Acknowledgements
  9. Supporting Information

During this study, we have refined our pharmacophore model collection for 11β-HSD inhibitors. Additionally, we have identified new 11β-HSD inhibitors from already used drugs, natural products, endocrine disrupting chemicals, and synthetic chemicals. All the refined pharmacophore models showed improved performance and quality and will be used for further virtual screening purposes.

Acknowledgements

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Methods
  5. 3 Results
  6. 4 Discussion
  7. 5 Conclusions
  8. Acknowledgements
  9. Supporting Information

Anna Vuorinen and Daniela Schuster thank the University of Innsbruck, Nachwuchsförderung for financing this work. Daniela Schuster is financed by the Erika Cremer Habilitation Program from the University of Innsbruck. This project was supported by the Tyrolean Science Fund (TWF) and the Austrian Academy of Science (ÖAW).

Supporting Information

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Methods
  5. 3 Results
  6. 4 Discussion
  7. 5 Conclusions
  8. Acknowledgements
  9. Supporting Information

As a service to our authors and readers, this journal provides supporting information supplied by the authors. Such materials are peer reviewed and may be re-organized for online delivery, but are not copy-edited or typeset. Technical support issues arising from supporting information (other than missing files) should be addressed to the authors.

FilenameFormatSizeDescription
minf_201300063_sm_miscellaneous_information.pdf2505Kmiscellaneous_information

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.