European Journal of Medicinal Chemistrycomponents in Pipeline Pilot 8.5 [44]. For the prospective screening, the Maybridge_HitDiscover database was downloaded from the Maybridge homepage

lable at ScienceDirect

European Journal of Medicinal Chemistry 124 (2016) 49e62

Contents lists avai

European Journal of Medicinal Chemistry

journal homepage: http: / /www.elsevier .com/locate/ejmech

Research paper

Evaluation of selected 3D virtual screening tools for the prospectiveidentification of peroxisome proliferator-activated receptor (PPAR) gpartial agonists

T. Kaserer a, 1, V. Obermoser b, 1, A. Weninger b, R. Gust b, D. Schuster a, *

a Computer-Aided Molecular Design Group, Institute of Pharmacy/Pharmaceutical Chemistry and Center for Molecular Biosciences Innsbruck (CMBI),University of Innsbruck, Innrain 80-82, 6020 Innsbruck, Austriab Pharmaceutical Chemistry and Center for Molecular Biosciences Innsbruck (CMBI), University of Innsbruck, Innrain 80-82, 6020 Innsbruck, Austria

a r t i c l e i n f o

Article history:Received 3 March 2016Received in revised form14 July 2016Accepted 28 July 2016Available online 6 August 2016

Keywords:Method comparisonDockingPharmacophore modelingShape-based modeling2D-similarity based searchPeroxisome proliferator-activated gammapartial agonists

Abbreviations: A, anion; Acc, accuracy; Ar, aromaticcurve; C, cation; DMEM, Dulbecco's Modified Eagle MEF, enrichment factor; FN, false negatives; FP, falseresonance energy transfer; GST, gluthatione S-transfeHBA, hydrogen bond acceptor; HBD, hydrogen bonenrichment factor; maxFN, maximum false negativepositive rate; maxTN, maximum true negative rate; mrate; MI, metal interaction; MNA, Multilevel Neighbtively ionizable; OE, overall enrichment; PASS, PredSubstances; PI, positively ionizable; PPAR, peroxisoceptor; R, ring feature; SD, standard deviation; SEA, SiTc, Tanimoto coefficient; TN, true negatives; TP, truedione; XVol, exclusion volume.* Corresponding author.

E-mail address: [email protected] (D. Sc1 These authors contributed equally to the publicat

http://dx.doi.org/10.1016/j.ejmech.2016.07.0720223-5234/© 2016 The Author(s). Published by Elsev

a b s t r a c t

The peroxisome proliferator-activated receptor (PPAR) g regulates the expression of genes involved inadipogenesis, lipid homeostasis, and glucose metabolism, making it a valuable drug target. However, fullactivation of the nuclear receptor is associated with unwanted side effects. Research therefore focuses onthe discovery of novel partial agonists, which show a distinct protein-ligand interaction patterncompared to full agonists. Within this study, we employed pharmacophore- and shape-based virtualscreening and docking independently and in parallel for the identification of novel PPARg ligands. Theten top-ranked hits retrieved with every method were further investigated with external in silicobioactivity profiling tools. Subsequent biological testing not only confirmed the binding of nine out of the29 selected test compounds, but enabled the direct comparison of the method performances in a pro-spective manner. Although all three methods successfully identified novel ligands, they varied in thenumbers of active compounds ranked among the top-ten in the virtual hit list. In addition, these com-pounds were in most cases exclusively predicted as active by the method which initially identified them.This suggests, that the applied programs and methods are highly complementary and cover a distinctchemical space of PPARg ligands. Further analyses revealed that eight out of the nine active moleculesrepresent novel chemical scaffolds for PPARg, which can serve as promising starting points for furtherchemical optimization. In addition, two novel compounds, identified with docking, proved to be partialagonists in the experimental testing.© 2016 The Author(s). Published by Elsevier Masson SAS. This is an open access article under the CC BY

license (http://creativecommons.org/licenses/by/4.0/).

feature; AUC, area under theedium; EE, early enrichment;positives; FRET, fluorescencerase; H, hydrophobic feature;d donor; maxEF, maximumrate; maxFP, maximum falseaxTP, maximum true negativeorhoods of Atoms; NI, nega-iction of Activity Spectra forme proliferator-activated re-milarity Ensemble Approach;positives; TZD, thiazolidine-

huster).ion.

ier Masson SAS. This is an open ac

1. Introduction

The peroxisome proliferator-activated receptors (PPARs) belongto the class of nuclear receptors. Upon activation, the receptorforms heterodimers with the retinoid X receptor (RXR) to controlthe expression of its targets genes [1]. PPARg is highly expressed inadipose tissue [2], where it is involved in adipogenesis [3], lipidhomeostasis, and glucose metabolism (reviewed in Ref. [4]). It getsactivated by a variety of endogenous compounds such as eicosanoid15-deoxy-D12,14-prostaglandin J2 or the fatty acid arachidonic acid[5], and by a number of synthetic [6e11] and natural compounds[12]. Most prominently, PPARg was identified as the moleculartarget of the thiazolidinedione (TZD) class of antidiabetic drugs[7,8], including the blockbuster drugs rosiglitazone and pioglita-zone. TZDs, of which most are full agonists, promote increased in-sulin sensitivity. However, their administration has also been

cess article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e6250

associated with severe side effects such as gain of body weight andcongestive heart disease [13]. These deleterious effects were foundto be diminished whereas beneficial effects on insulin sensitivitywere maintained, when PPARg was only partially activated [6,14].The design of several partial agonists of PPARg with improved sideeffect profiles have been reported [11,15,16], including the selectiveangiotensin receptor blocker telmisartan [9] and its analogues [17].Partial agonists were shown to display a different interaction modein the ligand-binding domain compared to full agonists. Mostprominently, other than full agonists, partial agonists do not sta-bilize helix H12 via hydrogen bonding with Tyr473 [18].

Virtual screening tools are nowadays well integrated in the drugdevelopment process to complement and support experimentalhigh-throughput screenings in the selection of the most promisingdrug candidates [19]. Several different virtual screening methodsare available for this purpose [20] (please refer to www.click2drug.org for a comprehensive list of virtual screening tools), and many ofthem have already been successfully applied for the identificationof novel PPARg ligands [21e28]. However, whether there are dif-ferences in their performances, and if so, which one is the mostsuitable for addressing nuclear receptor-related issues, and PPAR-related in particular, is still unclear. Several method comparisonshave been published throughout the last years [29e38], whichpoints out the raising interest in this topic. Unfortunately, thedifferent methodical set-ups of these studies hamper their com-parison on a larger scale. Therefore, a comprehensive and pro-spective evaluation of the same methods employing identicaldatasets still remains to be accomplished.

We have already investigated the performances of selectedcommon virtual screening tools for the identification of novelbioactive molecules for cyclooxygenases-1 and -2 as representa-tives of classical enzymes [39], and for members of the cytochromeP450 (CYP) superfamily involved in metabolic turnover [40].Intriguingly, we could observe quite distinct performances of thetools, suggesting that different tools might be better suited to meetthe requirements of the various target classes. Nuclear receptorsdisplay different properties concerning the structure of the proteinsand the ligands compared to the two examples investigated so far,which might be attributed to their different biological functions.Therefore we assumed that our findings so far may not be extrap-olated to nuclear receptors. To investigate the advantages andlimits of selected common virtual screening tools also for this targetclass, we selected PPARg as a case study representing nuclear re-ceptors and applied the same study design in linewith our previousinvestigations [39]. As mentioned above, research efforts con-cerning PPARg shifted towards the identification of partial agonistsrather than full agonists. Besides the identification of the mostsuitable virtual screening method for the identification of novelPPARg ligands, we therefore additionally aimed to investigate theability of the applied methods to reflect this specific binding modethat results in the partial activation of the receptor.

2. Methods

2.1. Study design

Analogous to our previous study [39], we generated PPARgpartial agonist pharmacophore- and shape-based models, andestablished a docking protocol that could discriminate betweenPPARg partial agonists, PPARg full agonists, and inactive com-pounds. All optimized models and the docking workflowwere usedfor virtual screening of the commercial Maybridge database (www.maybridge.com). The ten top-ranked hits from each of the threemethods were selected for further investigations andmerged to the“overall hit list”. In the next step, we analyzed whether these

compounds were also predicted by the other two virtual screeningmethods above the activity cut-off defined during the model gen-eration and theoretical validation. In addition, all compounds wereindependently and in parallel investigated with external 2D- and3D bioactivity profiling tools. All generated in silico predictionswere summarized in a prediction matrix. After biological testing,the performances of all the applied tools were evaluated andcompared. The workflow is depicted in Fig. 1.

2.2. Hardware specification

All processes and predictions were performed on a multi-coreworkstation with 2.4 þ GHz, 8 GB of RAM, a 1 þ TB fast massstorage, and a NVIDIA graphical processing unit. All programs runon the Windows 7 platform.

2.3. Datasets

Known active partial agonists of PPARg were manually assem-bled from the literature. Only compounds that activated the re-ceptor from 15% up to a maximum of 80% compared to the fullagonist control in a transactivation assay, and where direct bindingwas either shown by scintillation proximity assay or crystallo-graphic datawere included. In total, 51 knownpartial agonists wereincluded in the “partial agonist” dataset (a detailed list is providedin Table S1 of the supporting information).

To investigate whether the generated models and the dockingworkflow can discriminate between full and partial agonists, also a“full agonist” dataset was included. This dataset contained 14known full agonists from the literature for which direct bindingwas shown either via X-ray crystallography or a scintillationproximity assay and that activated the receptor >80% in a trans-activation assay (for a detailed list of compounds please refer toTable S2 of the supporting information). Some of these compoundsoriginated from the same chemical series as known partial agonistsand were therefore especially useful for investigating the structuralcharacteristics determining the biological activity.

PPARg was included as target in the ToxCast dataset and eval-uated in a fluorescence polarization assay [41]. The CAS-numbers ofall compounds that were inactive in this dataset against PPARgwere extracted and converted to a sd-file using Pipeline Pilotversion 9.1 [42] script “Search PubChem for CAS Number”. Thereby,a database of 799 unique structures (Table S3 in the supportinginformation) was generated. A detailed description of this protocolis provided in Fig. S1 of the supporting information. In addition, 13compounds which proved to be inactive in in vitro binding assayswere manually assembled from the literature (a detailed list isprovided in Table S4 of the supporting information). In total, thisled to an “inactives” dataset containing 812 known inactivecompounds.

A cdx-file was created for all literature-derived compounds us-ing ChemBioDraw Ultra 11.0 [43]. These cdx-files were then con-verted to sd-files using the ChemDraw Reader and SD Writercomponents in Pipeline Pilot 8.5 [44].

For the prospective screening, the Maybridge_HitDiscoverdatabase was downloaded from the Maybridge homepage (www.maybridge.com, access date 27 February 2014). This compoundcollection contained about 52,000 diverse molecules that arecommercially available. According to the information on thehomepage of the provider, the majority of compounds does fulfillgenerally acknowledged criteria of drug-likeness (http://www.maybridge.com/portal/alias__Rainbow/lang__en/tabID__146/DesktopDefault.aspx). As a consequence, we did not apply addi-tional pre-screening filters, which would have introduced addi-tional bias to this study: The application of an additional filter

Fig. 1. Study design.

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e62 51

would have had an impact on the composition of the hit list andaccordingly change the compounds selected for testing. Therefore,the performance of the virtual screening tools would depend notonly on the intrinsic properties of the method, but also on thecalculated ADMET properties of the screening database.

In principle, multiple commercially available compound col-lections can be employed for virtual screening. The composition ofthese databases may, however, also have an impact on the results.We therefore applied just one unmodified screening database toallow for the comparison of the methods performances withoutany influence of the data set and its predicted pharmacokineticproperties.

2.4. Prospective virtual screening tools

Analogous to our previous studies [39,40], LigandScout 3.1 [45]was employed for pharmacophore-based screening, vROCS 3.0.0[29,46] for shape-based screening, and GOLD [47,48] version 5.2was applied for all docking studies. A detailed description of theprograms and the virtual screening parameters is provided inSection S1 of the supporting information. The virtual hits retrievedfrom prospective screening were ranked according to the respec-tive fitness scores and the top-ranked molecules were selected forbiological testing. Whenever close analogues were ranked in thetop-ten, the one with the lower score was discarded and the next-ranked molecule was selected. Also compounds with reduced sta-bility such as esters were discarded.

2.5. External profiling tools

Several commercial and open source bioactivity profiling toolsare available and analogous to our previous studies [39,40] we alsoinvestigated whether these tools can correctly predict the activityof selected compounds with regard to a specific target. In detail,predictions were generated with the 2D similarity-based programsSEA [49] and PASS [50] and the pharmacophore-based profilingtools PharmMapper [51]. These tools are readily available onlineand may be very valuable because of their simple and fast appli-cation. A detailed description of these tools is provided in SectionS2 of the supporting information.

2.6. Biological testing

Pioglitazone and telmisartan were extracted as previouslydescribed [17]. The identity and purity of both compounds wasconfirmed via 1H-NMR prior to their usage. Identity and purity dataof the commercial test compounds were provided by Maybridge.Compounds were re-analyzed via HPLC whenever data weremissing. Detailed information regarding HPLC analyses is providedin Section S3 in the supporting information. For all active com-pounds, the identity and a purity of �95% was confirmed.

2.7. Time-resolved fluorescence resonance energy transfer (FRET)in vitro

A cell-free LanthaScreen TR-FRET PPARg competitive bindingassay kit [52] (Life Technologies, Darmstadt, Germany) was usedaccording to the manufacturer's protocol. Briefly, the human re-combinant ligand binding domain of PPARg is tagged with glu-thatione S-transferase (GST) and can be recognized by a terbium-labelled anti-GST antibody. Dilution series of the tested com-pounds in the indicated concentrations were produced in tripli-cates and incubated with the PPARg-LBD, the anti-GST antibody,and Fluoromone™ (5 nM), a fluorescein-labelled pan-PPAR agonist.The maximum concentration of pioglitazone and telmisartan rep-resented the positive assay controls and DMSO (FLUKA, Buchs,Switzerland) served as the negative control. Binding of thefluorescein-labelled agonist and thus resulting close proximity ofterbium and fluorescein increased the FRET signal, competitivedisplacement of the fluorescein-labelled agonist by tested com-pounds decreased the FRET signal. In case of autofluorescence ofthe tested compound, the fluorescence intensity of compound atthe given concentration was measured as blank and subtractedfrom the TR-FRET signal. Fluorescence intensities were measuredwith a filter-based Enspire plate reader (Perkin Elmer) and the FRETsignal was calculated by the ratio of the emission signal at 520 nm(fluorescein) and 495 nm (terbium), respectively. Instrument set-tings were conform to the suggestions of the manufacturer ofLanthaScreen assays and were verified with a LanthaScreen Tb in-strument control kit (Life Technologies, Darmstadt, Germany).

All compounds that showed at least 40% of the activity of thepositive control pioglitazone were considered as active.Concentration-response curves were determined for the mostactive molecules, which showed at least 60% of the pioglitazoneactivity.

2.8. Transactivation assay

The transactivation assaywas conducted as described elsewhere[17]. COS-7 cells (AATC) were seeded in 96-well plates at a densityof 104 cells/well in Dulbecco's Modified Eagle Medium (DMEM)with 4,5 g/L glucose, 584mg/L L-glutamine, supplementedwith 10%FCS (Biochrom, Berlin, Germany) and 1% penicillin/streptomycin(Life Technologies, Darmstadt, Germany), and incubated at 37 �Cand 5% CO2 for 24 h prior to transfection. After changing to serum-and antibiotic-free medium, cells were transiently transfected withpGal5-TK-Luc (30 ng), pGal4-PPARg-LBD (3 ng) and pRenilla-CMV(1 ng) in OptiMEM (25 mL; Gibco, Darmstadt, Germany). After 4 h,the selected compounds, pioglitazone, or vehicle (DMSO fromFLUKA, Buchs, Switzerland) were added at indicated concentra-tions, and luciferase activity was measured after 42 h. Transienttransfection (Lipofectamin 2000; Life Technologies, Darmstadt,Germany) and dual-luciferase reporter assay (Promega, Germany)were performed according to the manufacturer's protocol.

All compounds, the controls pioglitazone and telmisartan, andDMSO were tested in triplicates at a concentration of 10 mM in a

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e6252

first screening run. The screening was repeated once and the meanpercent activation compared to the control pioglitazone and the %standard deviation (SD) were calculated. For compounds with amean e SD higher than the mean of the DMSO control þ SD (12.1%receptor activation at 10 mM), the EC50 value and the maximum % ofreceptor activation compared to pioglitazone were determined.

2.9. Analysis of the results

During the theoretical validation, the enrichment factor (EF)[53] was calculated according to Equation (1):

EF ¼ TP=nA=N

(1)

In this equation, TP represents the number of true positives (i.e.active virtual hits), n the number of all virtual hits, A the number ofall active molecules in the validation dataset, and N the number ofall compounds in the validation dataset. This metric, however, ishighly dependent on the relation of active and inactivemolecules inthe dataset. To provide more independent quality metrics, wenormalized the EF to themaximum (max) EF and calculated the % ofthe maxEF a model yielded. In addition, the area under the curve(AUC) of the receiver operating characteristic (ROC)-plot wascalculated as described by Triballeau et al. [54].

After the biological testing, the performances of the prospectivevirtual screening tools were analyzed in three categories.

In the first category, early enrichment (EE), we investigated,which percentage of the top-ten ranked compounds was active inthe biological testing (Equation (2)).

EE ¼ number of active compounds in the top� 10 hit list10

�100(2)

This quality metric was only calculated for the three methodspharmacophore modeling, shape-based modeling, and docking,because the external profiling tools do not allow for virtualscreening of large compound databases. Using these tools, onecompound after the other has to be profiled, which prohibits theinvestigation of large compound libraries as accomplished with theother three methods. All other quality parameters described belowwere also calculated for the bioactivity profiling tools.

In the second category, overall enrichment (OE, Equation (3)),we investigated whether the compounds were also predicted bythe other methods above the activity-cut-off. For ROCS and dock-ing, the activity cut-offs were defined during the theoretical vali-dation. For SEA, a cut-off of E-value � �4 was selected, as proposedby Lounkine et al. [55]. For PASS, a cut-off of Pa � 0.5 was applied,because most of the active compounds were proven to be distrib-uted above this value [56].

OE¼ number of predicted and active compoundsnumber of predicted compounds from themerged hit list

�100

(3)

In the last category, we analyzed how many of the predictionswere correct. A prediction was considered as correct, when acompound was predicted to be active and indeed was active in thebiological assay, but also when a compoundwas not predicted to beactive and was inactive in the experimental testing. We refer to thisvalue as accuracy (Acc [57], Equation (4)). For this purpose, thenumber of TP, false positive (percentage of predicted but biologi-cally inactive compounds, FP), true negative (percentage of com-pounds that were not predicted and indeed were inactive in thebiological testing, TN), and false negative (number of compounds

that were not predicted, but were active in the biological testing,FN) hits was calculated.

Acc ¼ TP þ TNTP þ FP þ TN þ FN

*100 (4)

Since the TP, TN, FP, and FN rates heavily depend on the actualcomposition of active and inactive molecules in the overall hit list,these values were also normalized. For example, the normalized TPrate was calculated as the percentage of the maxTP rate retrieved.For the maxTP rate, the number of actual active hits was dividedthrough the number of overall compounds in the overall hit list. ThemaxTN, maxFP, and maxFN rates were calculated alike. Thesevalues are provided in addition to better assess the performances ofthe applied methods.

3. Results

3.1. Pharmacophore modeling

In total, 10 models using five PDB entries (2Q5S [18], 3H0A [58],2Q5P [18], 3FUR [11], and 3QT0 [59]), and one additional ligand-based model were generated with LigandScout 3.1 [45]. For theprospective part, however, three optimized models were selected,because they performed very well in the theoretical validation andtogether found the majority of compounds in the “partial agonist”dataset.

The first selected model was generated with the crystal struc-ture of PPARg in complex with the partial agonist nTZDpa (PDB-entry 2Q5S [18]). The refined model pm-2q5s (Fig. 2A) mapped 37out of the 51 (72.5%) partial agonists in the dataset, 64 inactivecompounds out of 812 (7.9% of the dataset), but no full agonist. Ittherefore retrieved an EF of 6.2, representing 37% of the maxEF.

The second model was based on the crystal structure of tetra-hydronaphthalene derivative 1 (Fig. 3) in complex with PPARg(PDB-entry 3H0A [58]). The optimized model pm-3h0a (Fig. 2B)matched only 9 out of 51 (17.6%) partial agonists, but also only 7 outof 812 (0.9%) inactive compounds in addition to one out of 14 fullagonists (7.1%). It yielded an EF of 9.0 (53% of the maxEF).

The last model selected for the prospective screening was aligand-basedmodel generated with the known partial agonists GQ-16 2 [15] and PA-082 3 [60] (Fig. 3). For these two molecules, Kivalues of 0.16 mM [15] and 0.8 mM [60] were determined, and theyactivated the receptor up to approximately one third [15] and 40%[60] compared to the full agonist rosiglitazone, respectively. Thefinalized pharmacophore model, pm-PPARg-pAg (Fig. 2C), found 9out of the 51 (17.6%) partial agonists and 7 out of the 812 (0.9%)inactive compounds. It did not map any full agonist of the dataset.This led to an EF of 9.5, which represents 56% of the maxEF.

In combination, all three models covered 48 out of the 51 partialagonists (94.1%), but only retrieved one out of the 14 full agonists(7.1%) and 74 out of the 812 (9.1%) inactive molecules. Together,they yielded an EF of 6.5, corresponding to 38% of the maxEF.Finally, they retrieved an AUC of 0.92.

In the prospective screening of the commercial Maybridgedatabase (52,000 entries), 9231 unique compounds mapped atleast one of the models. In detail, model pm-2q5s mapped 7393molecules, model pm-3h0a retrieved 177 virtual hits, and modelpm-PPARg-pAg matched 2369 compounds. All virtual hits wereranked according to their relative geometric pharmacophore fitscore, and the top ten ranked diverse and chemically stable mole-cules were selected for biological testing. A detailed list of theselected compounds and their highest relative Fit values is pro-vided in Table 1.

Fig. 2. Pharmacophore models for PPARg partial agonists. (A) Model pm-2q5s was based on the PDB-entry 2Q5S [18]. It consisted of 4 hydrophobic (H) features, 2 hydrogen bondacceptors (HBAs) to Ser342 and, water-mediated, to Arg288 and Glu343. Tyr473 of helix H12 is not involved in the binding and far away. (B) Model pm-3h0a was created using thePDB-entry 3H0A [58]. It contained three Hs, one negatively ionizable feature (NI), and one HBA. The latter two represented the interaction with Arg288. In addition, the modelcontained four optional HBAs. Partial agonists do not have to match these features, but they were crucial for proper feature alignment in LigandScout during virtual screening. (C)Model pm-PPARg-pAg was generated with the two known partial agonists GQ-16 2 (gray) and PA-082 3 (blue). It consisted of 2 H, two HBAs, and one aromatic ring (Ar) feature.Yellow spheres, H; red arrows and red spheres, HBAs; blue ring, Ar feature. (For interpretation of the references to colour in this figure legend, the reader is referred to the webversion of this article.)

Fig. 3. Structures of the training set compounds tetrahydronaphthalene derivative 1, GQ-16 2, PA-082 3, and isoxazolone derivative 4 [61].

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e62 53

3.2. Shape-based modeling

In the course of shape-based modeling, a total number of 50models was generated with vROCS 3.0.0 [29,46]; however, similarto pharmacophore modeling, only a selection of the best-performing models during the theoretical validation was appliedfor the prospective part.

The co-crystallized ligands of PPARg-compound complexesserved as query molecules for four shape-based models. These li-gands were preferentially selected for model generation, becausethey represent the biologically relevant conformations. Neverthe-less, also one model based on one low-energy conformation of theknown partial agonist isoxazolone derivative 4 [61] (Fig. 3) wasused for the successful retrieval of most of the compounds in the“partial agonist” dataset. The final shape-based models and theirperformances in the theoretical validation are summarized inTable 2. These models are depicted in Fig. 4AeE.

All models combined covered 35 out of the 51 (68.6%) partialagonists, but only mapped two out of 14 (14.3%) full agonists and 17

out of 812 (2.1%) known inactive molecules. Together, they yieldedan EF of 11.0%, representing 65% of the maxEF, and an AUC of 0.83.

In the prospective screening, 1848 unique compounds mappedat least one ROCS-model above the defined activity cut-off. Adetailed list of the number of virtual hits per model is provided inTable 2. For all virtual hits, the relative ComboScore was calculatedmanually as described in Section S1 in the supporting information,and this score was subsequently used to rank all mapping com-pounds. The ten top-ranked diverse and stable compounds wereselected for further in silico and experimental investigations. For adetailed list of selected compounds and their relative ComboScoresplease refer to Table 1.

3.3. Docking

In total, docking workflows were generated with GOLD [47,48]version 5.2 using the eight crystal structures 2Q5S [18], 2Q5P[18], 2Q6S [18], 2YFE [62], 3FUR [11], 3V9Y [63], 4A4V [64], and4A4W [64]. Finally, we selected only the best-performing docking

Table 1Prediction matrix for overall hit list. Values obtained in the top-ten hit lists of the respective methods are highlighted in bold.

Name LigandScouta ROCSb GOLDc SEAd PASSe PharmMapperf Activitym

Top-ranked pharmacophore modeling hits Compound 5 0.97h e e e e e e

Compound 6g 0.97h 1.215j e e e e e

Compound 7 0.97h e e e e e e

Compound 8 0.96h e e e e e e

Compound 9 0.96h e e e e e þCompound 10 0.96h e e e e e þCompound 11 0.96h e e e e e e

Compound 12 0.96h e e e e e þCompound 13 0.96h e e e e e e

Compound 14 0.95i e e e e e e

Top-ranked shape-based modeling hits Compound 15 e 1.265k e e e 0.607 þCompound 16 e 1.254j e e e e e

Compound 17 e 1.251j e e e e e

Compound 18 e 1.233l e e e e þCompound 19 0.93h 1.217l e e e e e

Compound 20 e 1.198j e e e e e

Compound 21 e 1.196l e e e e e

Compound 22 0.93h 1.192j 127.019 e e e e

Compound 23 e 1.189k e e e e e

Top-ranked docking hits Compound 24 e e 146.089 9.93e-4 e e þCompound 25 e e 144.178 e e 0.634 e

Compound 26 e e 141.653 e e e þCompound 27 e e 141.154 e e e e

Compound 28 e 1.011l 140.461 e e e e

Compound 29 e e 139.719 e e e þCompound 30 0.93h e 139.554 e e e e

Compound 31 e e 138.331 e e e e

Compound 32 e e 37.578 e e e þCompound 33 e e 136.966 e e e e

a Only highest relative pharmacophore fit score is listed for every compound, high values are desirable.b Only highest relative ComboScore is listed for every compound, high values are desirable.c Only highest GoldScore is listed for every compound, high values are desirable.d Only lowest E-value below the activity cut-off is listed for every compound, low values are desirable.e Only Pa values above activity cut-off are listed, high values are desirable.f Highest relative pharmacophore fit score retrieved with a model with at least 6 features, high values are desirable.g Consensus hit ranked in the top-ten of both the pharmacophore- and shape-based modeling hit list.h Identified with model pm-2q5s.i Identified with model pm-3h0a.j Identified with model shape-2q5s.k Identified with model shape-3vn2.l Identified with model shape-3fur.

m þ active in the biological testing, - inactive in the biological testing.

Table 2Shape-based models applied in this study.

Model Source ComboScorecut-off

No. of mapping partialagonists (%)

No. of mapping fullagonists (%)

No. of mappinginactives (%)

EF (%maxEF)

Hits in the prospectivescreening

Shape-2q5s 2Q5S [18] �1.20 14 out of 51 (27.5%) 1 out of 14 (7.1%) 5 out of 812 (0.6%) 11.8(70%)

423

Shape-2q5p 2Q5P [18] �0.94 15 out of 51 (29.4%) 1 out of 14 (7.1%) 3 out of 812 (0.4%) 13.4(79%)

247

Shape-3fur 3FUR [11] �1.20 5 out of 51 (9.8%) 0 out of 14 (0%) 4 out of 812 (0.5%) 9.4 (56%) 419Shape-3vn2 3VN2 [9] �0.88 9 out of 51 (17.6%) 1 out of 14 (7.1%) 7 out of 812 (0.9%) 9.0 (53%) 754Shape-

PPARg-pAgisoxazolone derivative4 [61]

�1.04 6 out of the 51 (11.8%) 0 out of 14 (0%) 0 out of 812 (0%) 16.9(100%)

81

Fig. 4. Shape-based models for PPARg partial agonists. The models were generated with (A) nTZDpa (PDB-entry 2Q5S [18]), (B) MRL24 (PDB-entry 2Q5P [18]), (C) INT131 (PDB-entry3FUR [11]), (D) telmisartan (PDB-entry 3VN2 [9]), and (E) one low-energy conformation of the known partial agonist isoxazolone derivative 4 [61]. Color features were added torefine the shape models: green sphere, ring feature; red sphere, anion; blue sphere, cation; yellow sphere, hydrophobic; red mesh, HBA. (For interpretation of the references tocolour in this figure legend, the reader is referred to the web version of this article.)

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e6254

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e62 55

protocol in the theoretical validation for the prospective part tolimit the time required for the docking of the about 52,000 mole-cules in the Maybridge database. The crystal structure of PPARg incomplex with the partial agonist amorfrutin B (PDB-entry 4A4W[64]) was employed for docking, because 18 out of the 51 (35.3%)partial agonists, but only 2 out of 14 (14.3%) full agonists and 4 outof 84 (4.8%) inactive compounds from the reduced “inactives”dataset for docking were scored above a GoldScore of 124.0. Sincethis GoldScore allowed for a good balance between the retrieval ofpartial agonists while excluding full agonists and inactive com-pounds, respectively, it was selected as activity cut-off. Thereby,docking yielded an EF of 2.2, representing 75% of the maxEF, and anAUC of 0.65. However, this comparably low AUC is rather due to thestrict activity cut-off selected, which consequently leads to asmaller portion of partial agonists ranked above this value, than to alarge number of full agonists or inactive compounds predicted to beactive.

In the prospective docking run, 809 unique compounds weredocked into the binding site of PPARg with a GoldScore of �124.0.All virtual hits were ranked according to the GoldScore and the top-ten ranked diverse and chemically stable molecules were selectedfor further investigations. For a detailed list of selected compoundsand their GoldScore, please refer to Table 1.

3.4. Generation of a prediction matrix

The top-ten ranked diverse and stable compounds retrievedwith the three methods pharmacophore modeling, shape-basedmodeling, and docking were merged to an overall hit list. Onecompound, compound 6, was ranked among the top-ten moleculesby both pharmacophore- and shape-based modeling. Therefore,the overall hit list contained 29 unique compounds. In the nextstep, it was investigated whether these compounds were alsopredicted by the other two methods above the defined activity cut-off in case they were not ranked among the top-ten. In addition, allcompounds were further profiled with the external bioactivityprofiling tools SEA, PASS, and PharmMapper (see below). All pre-dictions for every compound were then collected in a predictionmatrix (Table 1).

3.5. External profiling tools

All 29 compounds from the overall hit list were further inves-tigated with the external bioactivity profiling tools SEA [49], PASS[50], and PharmMapper [51]. SEA predicted one compound, com-pound 24, above the activity cut-off, while PASS did not calculate aPa value of�0.5 for a single compound. When using PharmMapper,two compounds, 15 and 25, mapped a PPARg pharmacophoremodel consisting of at least six features. All predictions generatedwith the three external bioactivity profiling tools for every com-pound were added to the prediction matrix (Table 1).

3.6. Biological results

In the TR FRET assay, displacement of the labelled ligand wasobserved for nine out of the 29 investigated compounds (Table 3and Fig. 5). These compounds therefore proved to bind to PPARg.For the five most active compounds, which showed at least 60% ofthe binding of pioglitazone, concentration-response curves weredetermined (Fig. 6). Compounds 10, 26, and 32 showed a sigmoidalcurve and a maximum ligand displacement of 50, 60, and 80%,respectively. For compound 26, a similar concentration-responsecurve as the positive control pioglitazone was observed, provingthat they have equal affinity towards PPARg (Fig. 6A). The tworemaining active compounds 15 and 24 showed linear

concentration-response relationships. The partial agonist telmi-sartan showed a higher affinity than the tested compounds andpioglitazone (Fig. 6).

To further elucidate the biological effects of the compounds, allof them were investigated in a transactivation assay. In this assay,however, only two of the molecules could induce higher geneexpression than the mean ± SD of the DMSO control. These resultsmay either be caused by the compounds pharmacokinetic proper-ties (low membrane permeability, low solubility in the cellularbuffer) or the biological activity (compounds are antagonists). Thefirst molecule, compound 24, activated PPARg up to 20.6± 3.9% at aconcentration of 10 mM. Compound 26, the second active com-pound, activated PPARg up to 29.8± 6.3% at a concentration of10 mM. For these two compounds, which also were the most activeones in the FRET assay, and for pioglitazone, concentration-response curves were also determined in the transactivationassay. These curves were sigmoidal (Fig. 7) and therefore allowedfor the calculation of EC50 values of 3.4 ± 0.9 mM and 21.7 ± 6.3 mMfor compound 24 and 26, respectively. An EC50 value of 0.5 ± 0.1 mMand 4.7 ± 0.3 mM was determined for the positive controls piogli-tazone and telmisartan, respectively. Pioglitazone is a full agonist,and its maximum receptor activationwas therefore set to 100%. Thetwo novel compounds 24 and 26 induced a maximum receptoractivation of 23.7 ± 2.2% and 41.9± 2.7% compared to pioglitazone.Similar to telmisartan, which showed a maximum receptor acti-vation of 48.2± 3.0% compared to pioglitazone, they activated thereceptor to a lesser extent and proved to be partial agonists in thetransactivation assay. The results of the biological testing for allactive compounds and the positive controls pioglitazone and tel-misartan are summarized in Table 3.

The structures of the inactive compounds as well as their bio-logical activity in the experimental testing are provided in Table S5in the supporting information.

3.7. Analysis of the applied methods

The performances of all applied tools and for every qualitymetric, as described in more detail in the Methods Section, arelisted in Table 4. No values for EE are included for the bioactivityprofiling tools SEA, PASS, and PharmMapper, because no testcompounds were selected based on predictions generated by them.A detailed graphical representation of all method performancesincluding EE, OE, Acc, TP, TN, FP, and FN rates is displayed in Fig. 8.An analysis of the relative quality metrics maxTP, maxTN, maxFP,and maxFN is provided in Fig. S2 of the supporting information.

4. Discussion

4.1. Evaluation of the applied programs

The main aim of this study was to evaluate the performances ofcommonly used, selected virtual screening tools such as thepharmacophore-based software LigandScout, the shape-basedmodeling program ROCS, and the docking software GOLD. Theterms “program” and “method” were applied exchangeable in thissection, because the selected programs were compared as repre-sentatives of the different methods. In addition, the ability of thefreely available bioactivity profiling tools SEA, PASS, and Pharm-Mapper to correctly classify the molecular target PPARg for the 29molecules of interest was investigated. The latter tools are readilyavailable online andmay thus allow for the fast and straightforwardidentification of novel targets for the compounds of interest.Despite the limited scope of this study - only ten molecules perprospective screeningmethodwere subjected to biological testing -we could already observe considerable differences in the

Fig. 5. Structures of the active compounds 9, 10, 12, 15, 18, 24, 26, 29, and 32.

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e6256

performances of the applied tools. The results reported in this studyshould support the application of the most suitable virtualscreening tool, whenever nuclear receptors are investigated withcomputational models in future studies.

Four of the novel PPARg ligands have been identified withdocking, proving it to be the most successful method in retrievingactive molecules in this study. This result is somewhat unexpected,as in our previous studies, docking was the method that retrievedfewer active compounds compared to the other two techniques(nevertheless, it had other advantages) [39,40]. In addition, two ofthe active compounds were ranked on first and third place. Theperformance of docking is often assumed to suffer from scoring[65], i.e. the correct ranking of compounds based on their actualactivity, and also we had observed, that all of the compounds thatwere correctly classified by docking in our previous studies werenot ranked in the top-two positions [39,40]. However, this study onPPARg ligands provides an example proving that prospectivedocking and scoring can be very successful if the scoring function isappropriate for the target. In our recent study on CYP metabolicenzymes, no theoretically validated docking workflows could begenerated for CYP2C9 and 3A4 [40]. We suggested that this findingis due to the larger binding pockets of these two targets (~470 Å3

and up to 2000 Å3 for CYP2C9 [66] and 3A4 [67], respectively),which offer more possibilities to incorrectly fit inactive compoundsinto the binding pocket and retrieve high-scoring poses [40].However, the binding pocket of PPARg was estimated ~1300 Å3

[68], which is much larger as e.g. the CYP2C9 binding site. Incontrast to the CYP enzymes, where the binding pockets are formedby one large cavity, the PPARg binding site consists of three Y-shaped narrow channels. This clear separation and the narrow

composition of the three channels facilitated a precise definition ofthe intended binding site, despite the large size of the overallbinding pocket. These results suggest that not only the size, but alsothe overall composition of the binding pocket influences thedocking performance.

Besides correct ranking and the identification of novel bioactivemolecules, docking proved to be highly suitable to distinguish be-tween full and partial agonists and represent the specific partialagonist binding mode. This docking workflow may therefore be avaluable tool for further studies.

The pharmacophore-based virtual screening protocol was suc-cessful in the identification of novel PPARg ligands. The exact ac-tivity, i.e. whether the compounds display agonistic or antagonisticactivity, could not be determined. Accordingly, we cannot evaluatewhether the pharmacophore models were also successful in dis-tinguishing between the distinct activity classes. All active com-pounds were identified with model pm-2q5s. This model was lessspecific in the theoretical validation, as it also mapped more than7% of the inactive molecules. Also in the prospective screening, itretrieved a large proportion of the database compared to the othermodels. Nevertheless, it proved a very valuable tool for rankingactive compounds at the top of the hit list and may therefore beapplied also in future PPARg-associated drug discovery projects.

According to the good performance during the theoretical vali-dation, the ROCS models also performed well in the prospectivevirtual screening. Although they did not identify as many novelactive compounds as docking and pharmacophore modeling,shape-based modeling contributed two novel PPARg ligands to theoverall hit list that would have been missed with the other twoprospective screening techniques. These two molecules,

Fig. 6. Concentration-response curves of compounds 24 and 26 (A) and compounds10, 15, and 32 (B), and the positive controls pioglitazone and telmisartan (A, B) for theFRET assay, respectively.

Fig. 7. Concentration-response curves of compounds 24 and 26 and the positivecontrols pioglitazone and telmisartan for the transactivation assay.

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e62 57

compounds 15 and 18, were in addition discovered with twodifferent models. In general, if one compares the five final models(Fig. 4), it is intriguing to observe how diverse they are. The shapesof model shape-2q5s (Fig. 4A) and shape-3fur (Fig. 4C) are almostcomplementary. This suggests that the geometry of the ligandsmaynot be such an important factor in PPARg binding. Nevertheless,ROCS proved to be successful in the retrieval of active compounds,and, similar to the other prospective in silico workflows developedin this study, represents a valuable tool for future projects.

In total, nine novel PPARg ligands were identified in this study.The most active ones of these compounds bind PPARg in the lowmicromolar range (EC50s of 2.1e5.4 mM) in the cell-free assaysimilar to the positive control pioglitazone (EC50 of 1.6 mM). Forcompound 24, an EC50 of 606 nMwas determined, thereby showingthat this compound has a higher binding affinity than pioglitazone.Although the remaining compounds are only weakly active, theidentification of nine novel active hits confirmed PPARg ligands canbe seen as remarkably successful. Only 29 compounds were testedat all, thereby leading to a hit rate of 31%. In the context of virtual hitactivity, Scior et al. [69] described too high expectations concerningthe retrieval of highly active molecules as the first pitfall in virtualscreening. The authors state that although the identification of high

affinity ligands would be desirable, the majority of initial hits areonly weakly active. Similar to initial hits derived from experimentalhigh-throughput screening, these molecules have to be improvedin further optimization steps.

Intriguingly, while all of the three prospective tools identifiednovel compounds, they only predicted their own virtual hits asactive, but none of the active molecules discovered with the othertwo programs. Consequently, none of the tools was a universalapproach to identify most active compounds. To investigatewhether the three programs cover a distinct chemical space, weanalyzed the structural diversity of the nine novel PPARg ligandswith the “Calculate Diversity Metrics” tool implemented in Dis-covery Studio 4.0 [70]. For the comparison, we selected the 2Dsimilarity metric ECFP4 and the Tanimoto coefficient (Tc), whichranks from zero to one. The higher the Tc, the more similar are thecompared compounds. Conversely, the smaller the value, the moredissimilar the compounds are from a 2D structural point of view.For the nine active molecules, an average fingerprint similarity of0.11 was calculated, with a minimum and maximum distance valueof 0.25 and 0.03. The two most similar compounds among thenewly identified ligands retrieved a Tc of 0.25, thereby suggestingthat all the novel ligands are structurally very distinct from eachother. This study highlights the complementarity of the appliedvirtual screening tools with respect to the chemical space the novelPPARg occupy. Similar observations were also reported by us [71]and others [32] before.

All molecules of the overall hit list were further investigatedwith the external bioactivity profiling tools SEA, PASS, andPharmMapper. SEA predicted only compound 24 to be active,which could be confirmed in the biological testing. SEA failed toidentify all other eight novel PPARg ligands, however, it correctlyclassified all remaining 20 inactive molecules.

The second 2D similarity-based tool PASS did not predict asingle compound above the activity cut-off. Consequently, it failedto correctly identify all novel PPARg ligands. However, similar toSEA, it correctly predicted the inactivity of the remaining 20compounds.

In accordance with our previous results, the 2D similarity-based

Table 3Biological screening results of the active molecules.

Compound TR-FRET assay Transactivation assay

% Replacement of ligand ±SD(10 mM)

% Replacement compared topioglitazone (10 mM)

EC50 (mM) % Activation ± SD(10 mM)

EC50 (mM) Maximum activation compared to 10 mMpioglitazone (%)

Compound9

32.6 ± 4.2 48 ± 6 n. d.a 3.5 ± 0.9 n. d. n. d.

Compound10

54.0 ± 10.6 80 ± 16 2.1 ± 0.3 11.2 ± 8.4 n. d. n. d.

Compound12

30.9 ± 5.9 46 ± 9 n. d. 8.4 ± 3.4 n. d. n. d.

Compound15

40.7 ± 5.3 60 ± 8 3.1 ± 0.6 8.3 ± 1.5 n. d. n. d.

Compound18

30.4 ± 2.9 45 ± 4 n. d. 5.3 ± 4.0 n. d. n. d.

Compound24

67.1 ± 4.8 99 ± 7 0.6 ± 0.1 20.6 ± 3.9 3.4 ± 0.9 23.7 ± 2.2

Compound26

58.0 ± 1.9 86 ± 3 2.8 ± 0.4 29.8 ± 6.3 21.7 ± 6.3 41.9 ± 2.7

Compound29

28.3 ± 9.7 42 ± 14 n. d. 5.3 ± 1.1 n. d. n. d.

Compound32

51.2 ± 2.4 76 ± 3 5.4 ± 0.7 15.4 ± 4.6 n. d. n. d.

Pioglitazone 67.5 ± 10.6 100 1.6 ± 0.2 100 ± 6.2 0.5 ± 0.1 100Telmisartan 61.3 ± 3.5 91 ± 5 0.7 ± 0.01 44.8 ± 3.5 4.7 ± 0.3 48.2 ± 3.0DMSO 4.2 ± 4.8 6 ± 7 n. d. 9.6 ± 2.5 n. d. n. d.

Table 4Analysis of the method performances.

Program EE (%) OE (%) Acc (%) TP (%) maxTP (%) TN (%) maxTN (%) FP (%) maxFP (%) FN (%) maxFN (%)

LigandScout 30.0 23.1 44.8 10.3 33.3 34.5 50.0 34.5 50.0 20.7 66.7ROCS 20.0 18.2 44.8 6.9 22.2 37.9 55.0 31.0 45.0 24.1 77.8GOLD 40.0 36.4 58.6 13.8 44.4 44.8 65.0 24.1 35.0 17.2 55.6SEA n. d.a 100.0 72.4 3.5 11.1 69.0 100.0 0.0 0.0 27.6 88.9PASS n. d. 0.0 69.0 0.0 0.0 69.0 100.0 0.0 0.0 31.0 100.0PharmMapper n. d. 50.0 69.0 3.5 11.1 65.5 95.0 3.5 5.0 27.6 88.9

a n.d. - not determined.

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e6258

methods SEA and PASS proved to be very restrictive. In our recentCYP [40] study, we assumed that this might be explained by theunderlying screening concept: Both programs are machine-

Fig. 8. Detailed graphical representation of all method performances. (A) Shows the EE, OE,the external bioactivity profiling tools SEA, PASS, and PharmMapper, because no compouncomposition of the hit lists obtained with every method with respect to TP, TN, FP, and FN

learning tools that calculate the probability of the query moleculeto be active against a specific target based on the 2D structure of thequery and already known active molecules. As mentioned above,

and Acc values retrieved by all applied methods. * The EE could not be determined fords were selected for prospective testing based on their predictions. (B) Displays therates.

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e62 59

the novel active compounds retrieved very low similarity scoreswhen we compared them to each other. This suggests that theypossess a very different 2D structure from so far known ligands (adetailed investigation of the novelty of the scaffolds is providedbelow). This could explain why it was very challenging for thesetools to predict these novel scaffolds as active, as no similar knownactive compounds are available for comparison. That SEA couldcorrectly identify at least one of the novel compounds is thereforeremarkable, especially, as the comparison with the most similarknown active compound yielded a Tc of 0.39.

In accordance with our findings in the CYP study [40], we as-sume, that these tools do not predict a lot of compounds as active,but if they do so, it is often correct. This may especially account forSEA, which, similar to the CYP1A2 results [40], retrieved an OE rateof 100%. Therefore, we consider it very useful for cherry pickingapproaches.

The last external bioactivity profiling tool which we applied,PharmMapper, is largely independent from the structures ofalready known active compounds, because it is based on theinteraction patterns of known active compounds rather than their2D structure. This tool matched two compounds to PPARg phar-macophore models, of which one was active in the biologicaltesting. However, similar to the results of the other externalbioactivity profiling tools reported here, it did not match many ofthe compounds in general.

4.2. Novel compounds

To further investigate the novelty of the nine active compoundswith respect to known PPARg ligands, we compared their struc-tures to that of the known active molecules in the full and partialagonist dataset that we employed for the theoretical validation. Wedecided to include all active molecules in this comparison, becausefor seven out of the nine novel ligands their exact biological activitycould not be determined. For the calculation of the similarity, weapplied the “Find similar Molecules” tool implemented in Discov-ery Studio version 4.0 [70], again using ECFP4 to calculate the fin-gerprints. This tool also determines the similarity of compoundswith the Tc. In detail, the similarities between compounds 9, 10, 12,15, 18, 24, 26, 29, and 32 and the most similar compound in thedataset were calculated as 0.20, 0.28, 0.15, 0.23, 0.34, 0.30, 0.20,0.22, and 0.15, respectively. The novel molecules are therefore verydissimilar to the already known active compounds applied duringtheoretical validation. In addition, we searched for similar com-pounds in SciFinder. For all compounds, no reference related toPPARgwas available. In fact, for seven out of the ninemolecules, nota single referencewas deposited in SciFinder at all. In a next stepwe

Fig. 9. (A) Comparison of the binding poses of the full agonist pioglitazone (dark gray, PDB-enOnly pioglitazone occupies the side pocket formed by helix H12 and interacts with Tyr473. (crystallized partial agonist amorfrutin B (gray) in the PPARg crystal structure (PDB-entry 4A4away from Tyr473. (For interpretation of the references to colour in this figure legend, the

investigated, whether molecules with a similarity of�80 comparedto the novel ligands identified in this study were associated withPPARg before. SciFinder also employs the Tc for determining thesimilarity of compounds. However, in this case the values arenormalized to 100. The maximum value that can be retrieved, i.e.for identical molecules, is therefore 100 instead of one. For com-pound 15, one derivative (similarity score of 86) was related tohyperlipidemia and obesity, however, these effects were reportedto be caused by another target [72]. Analogues of compound 18were further linked to the treatment of hypercholesterolemia [73],LDL receptor expression [74] (similarity scores of 82), and PPARgmodulation [75] (similarity score of 80). All other compounds werenot associated with PPARg before. These results also highlight theability of the applied tools to retrieve structurally novel bioactivecompounds.

To investigate whether the novel PPARg ligands may be falsepositive hits that interfere with the assay, they were analyzed withthe PAINS filter provided by the endocrine disruptome homepage[76]. However, according to these predictions, the compounds arenot under risk for pan-assay interference. The application of furtherfilters is discussed in Section S4 in the supporting information.

4.3. Partial agonism

This study did not only evaluate the performances of the appliedtools with respect to the detection of novel PPARg ligands, but theidentification of novel partial agonists in particular. In principle,PPARg has a hydrophobic Y-shaped ligand-binding domain thatconsists of 13 a-helices and one b-sheet. The area containing helixH12 is part of the activation function 2 domain that forms the co-activator binding-site. Helix H12 is very flexible, and its stabiliza-tionwith small molecules induces co-activator recruitment and fullactivation of the receptor. Hydrogen bondingwith residues on helixH12 or in close vicinity, especially His323, His449, and Tyr473 wereobserved to be a common feature of many full agonists [18].

Partial agonists, however, seem to exert their effects via distinctprotein-ligand interactions: In contrast to full agonists, partial ag-onists occupy a different region around the b-sheet. In particular,partial agonists can form hydrogen bonds with Ser342, Arg288,and, water-mediated, with Glu343 [18]. In addition, extensive hy-drophobic interactions involving for example Ile281, Ile247, Leu255,and Gly285 have been observed. Partial agonists therefore do notstabilize helix H12, but rather helix H3 and the b-sheet. This leadsto altered conformational changes, co-regulator recruitment, andgene expression profiles [18]. Fig. 9A shows the distinct areas of thePPARg binding pocket the full agonist pioglitazone (PDB-entry2XKW [77]) and the partial agonist amorfrutin B (PDB-entry 4A4W

try 2XKW [77]) and the partial agonist amorfrutin B (light gray, PDB-entry 4A4W [64]).B) The docking poses of compounds 24 (pink) and 26 (blue) in comparison with the co-W [64]). All compounds are positioned in close proximity to Ser342 and Arg288, but farreader is referred to the web version of this article.)

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e6260

[64]) occupy.In the cellular assay, for two out of the nine compounds that

proved to bind to the PPARg ligand-binding domain, a biologicalactivity was observed. These two active compounds activate thereceptor up to 23.7% (p < 0.0001, One way ANOVA and Dunnett'sposthoc test compared to DMSO control) and 41.9% (p< 0.0001, Oneway ANOVA and Dunnett's posthoc test compared to DMSO con-trol), respectively. The EC50 of compound 26 is with 21.7 mM rela-tively high, however, one has to keep in mind that this compound isa first lead compound that did not undergo further chemical opti-mization steps so far. Due to their structural novelty, both mole-cules may represent interesting and promising starting points forfurther chemical refinement. Fig. 9B displays the docking poses ofcompounds 24 and 26. Both compounds are positioned in closeproximity to Ser342 and Arg288, but far away from Tyr473. Theseinteraction patterns were described for partial agonists before [18]and the retrieved docking poses are therefore in accordance withthe biological activity of compounds 24 and 26.

In detail, compound 24 was predicted to form hydrogen bondswith Arg288, Ser342, Cys285, Gly344, and Leu330. In the majorityof cases, the trifluoromethyl-moieties were involved in these in-teractions. Sulfonamide-oxygen atoms formed the remaininghydrogen bonds with Arg288 and Leu330. In addition, hydrophobiccontacts with residues Phe226, Leu228, Ile281, Phe287, Ala292,Leu333, Val339, Ile341, Met348, Leu353, Phe363, and Met364 werepredicted by the docking pose.

For compound 26, docking results suggested hydrogen bondswith Arg288, Leu340, andewater-mediated - with Ser342, Gly344,and Ile341. Besides, compound 26 was predicted to be involved inhydrophobic contacts with Pro227, Leu228, Ile281, Leu333, Val339,Ile341, Met348, and Leu353.

5. Conclusions

In summary, we applied selected common virtual screeningtools in parallel and independently for the identification of novelPPARg partial agonists. This parallel application allowed us todirectly compare the performances of all tools in a prospectivemanner. Similar to our previous studies, where we investigatedother drug target classes, we could observe substantial differencesin the performances of the tools: The three prospective programsLigandScout, ROCS, and GOLD proved to be very successful in theidentification of novel compounds, although we could observedifferences in the numbers of novel ligands they discovered. Thebest-performing tool in this study was GOLD, which ranked fouractive compoundswithin the top ten of the hit list. Intriguingly, twoof these molecules were ranked at positions one and three, therebyshowing the high suitability of our docking workflow for discov-ering novel PPARg ligands. In contrast, the external bioactivityprofiling tools SEA, PASS, and PharmMapper appeared to be veryrestrictive. This led to a very limited retrieval of the active com-pounds. However, also the numbers of false positive hits wereextremely low, making them very valuable tools for cherry picking.Interestingly, none of the active compounds identified within thisstudy was predicted as active by any other of the prospectivelyapplied programs (LigandScout, ROCS, and GOLD) except the onethat led to their initial identification. This finding suggests that theprograms are highly complementary in terms of the chemicalscaffolds they cover. This is further supported by our similarityanalysis, proving that the novel active compounds have diverse 2Dchemical structures.

In total, nine novel PPARg ligands out of 29 tested compounds(31.0%) were successfully identified within this study. The majorityof these compounds is structurally diverse from all so far knownPPARg ligands and may therefore represent promising starting

points for further chemical optimization. These optimization stepsmay not only include the target-related activity, but also pharma-cokinetic properties, as seven out of the nine compounds did notshow any activity in the cell-based assay. However, these resultsmake the two compounds 24 and 26, which proved to be partialagonists in the transactivation assay, interesting lead structures, astheir activity suggests an improved adverse event profile and adistinct protein-ligand interaction pattern compared to full ago-nists. Especially compound 26, which showed a sigmoidalconcentration-response curve and a similar binding affinity as thestandard compound pioglitazone in the FRET assay, should beinvestigated in further studies.

Acknowledgements

This project was supported by the foundation “Verein zurF€orderung der wissenschaftlichen Ausbildung und T€atigkeit vonSüdtirolern an der Landesuniversit€at Innsbruck“, the University ofInnsbruck (Young Talents Grant and a position within the ErikaCremer Habilitation Program), the Tyrolean Science Fund, and theAustrian Science Fund (P26782). We thank Peter Enoh for technicalassistance, Philipp Schuster for comments on the manuscript, andOpeneye and Inte:Ligand for providing software free of charge.

Appendix A. Supplementary data

Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.ejmech.2016.07.072.

References

[1] V. Chandra, P. Huang, Y. Hamuro, S. Raghuram, Y. Wang, T.P. Burris,F. Rastinejad, Structure of the intact PPAR-g-RXR-a nuclear receptor complexon DNA, Nature 456 (2008) 350e356.

[2] A.J. Vidal-Puig, R.V. Considine, M. Jimenez-Li~nan, A. Werman, W.J. Pories,J.F. Caro, J.S. Flier, Peroxisome proliferator-activated receptor gene expressionin human tissues. Effects of obesity, weight loss, and regulation by insulin andglucocorticoids, J. Clin. Invest. 99 (1997) 2416e2422.

[3] P. Tontonoz, E. Hu, B.M. Spiegelman, Stimulation of adipogenesis in fibroblastsby PPARg2, a lipid-activated transcription factor, Cell 79 (1994) 1147e1156.

[4] M. Lehrke, M.A. Lazar, The many faces of PPARg, Cell 123 (2005) 993e999.[5] S.A. Kliewer, S.S. Sundseth, S.A. Jones, P.J. Brown, G.B. Wisely, C.S. Koble,

P. Devchand, W. Wahli, T.M. Willson, J.M. Lenhard, J.M. Lehmann, Fatty acidsand eicosanoids regulate gene expression through direct interactions withperoxisome proliferator-activated receptors a andg, Proc. Natl. Acad. Sci. U. S.A. 94 (1997) 4318e4323.

[6] P.J. Larsen, K. Lykkegaard, L.K. Larsen, J. Fleckner, P. Sauerberg,K. Wassermann, E.M. Wulff, Dissociation of antihyperglycaemic and adverseeffects of partial perioxisome proliferator-activated receptor (PPAR-g) agonistbalaglitazone, Eur. J. Pharmacol. 596 (2008) 173e179.

[7] T.M. Willson, J.E. Cobb, D.J. Cowan, R.W. Wiethe, I.D. Correa, S.R. Prakash,K.D. Beck, L.B. Moore, S.A. Kliewer, J.M. Lehmann, The Structure�Activityrelationship between peroxisome proliferator-activated receptor g agonismand the antihyperglycemic activity of thiazolidinediones, J. Med. Chem. 39(1996) 665e668.

[8] J.M. Lehmann, L.B. Moore, T.A. Smith-Oliver, W.O. Wilkison, T.M. Willson,S.A. Kliewer, An antidiabetic thiazolidinedione is a high affinity ligand forperoxisome proliferator-activated receptor g (PPARg), J. Biol. Chem. 270(1995) 12953e12956.

[9] Y. Amano, T. Yamaguchi, K. Ohno, T. Niimi, M. Orita, H. Sakashita, M. Takeuchi,Structural basis for telmisartan-mediated partial activation of PPAR gamma,Hypertens. Res. 35 (2012) 715e719.

[10] M.-K. Kim, Y.N. Chae, S.-h. Choi, H.S. Moon, M.-H. Son, M.-H. Bae, H.-h. Choi,Y. Hur, E. Kim, Y.H. Park, C.S. Park, J.G. Kim, J.I. Lim, C.Y. Shin, PAM-1616, aselective peroxisome proliferator-activated receptor g modulator with pre-served anti-diabetic efficacy and reduced adverse effects, Eur. J. Pharmacol.650 (2011) 673e681.

[11] A. Motani, Z. Wang, J. Weiszmann, L.R. McGee, G. Lee, Q. Liu, J. Staunton,Z. Fang, H. Fuentes, M. Lindstrom, J. Liu, D.H.T. Biermann, J. Jaen, N.P.C. Walker,R.M. Learned, J.-L. Chen, Y. Li, INT131: a selective modulator of PPAR gamma,J. Mol. Biol. 386 (2009) 1301e1311.

[12] L. Wang, B. Waltenberger, E.-M. Pferschy-Wenzig, M. Blunder, X. Liu,C. Malainer, T. Blazevic, S. Schwaiger, J.M. Rollinger, E.H. Heiss, D. Schuster,B. Kopp, R. Bauer, H. Stuppner, V.M. Dirsch, A.G. Atanasov, Natural productagonists of peroxisome proliferator-activated receptor gamma (PPARg): a

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e62 61

review, Biochem. Pharmacol. 92 (2014) 73e89.[13] R.W. Nesto, D. Bell, R.O. Bonow, V. Fonseca, S.M. Grundy, E.S. Horton, M. Le

Winter, D. Porte, C.F. Semenkovich, S. Smith, L.H. Young, R. Kahn, Thiazolidi-nedione use, fluid retention, and congestive heart failure: a consensus state-ment from the american heart association and american diabetes association,Circulation 108 (2003) 2941e2948.

[14] C. Liu, T. Feng, N. Zhu, P. Liu, X. Han, M. Chen, X. Wang, N. Li, Y. Li, Y. Xu, S. Si,Identification of a novel selective agonist of PPARg with No promotion ofadipogenesis and less inhibition of osteoblastogenesis, Sci. Rep. 5 (2015)1e12.

[15] A.A. Amato, S. Rajagopalan, J.Z. Lin, B.M. Carvalho, A.C.M. Figueira, J. Lu,S.D. Ayers, M. Mottin, R.L. Silveira, P.C.T. Souza, R.H.V. Mour~ao, M.J.A. Saad,M. Togashi, L.A. Simeoni, D.S.P. Abdalla, M.S. Skaf, I. Polikparpov, M.C.A. Lima,S.L. Galdino, R.G. Brennan, J.D. Baxter, I.R. Pitta, P. Webb, K.J. Phillips,F.A.R. Neves, GQ-16, a novel peroxisome proliferator-activated receptor g(PPARg) ligand, promotes insulin sensitization without weight gain, J. Biol.Chem. 287 (2012) 28169e28179.

[16] J.P. Berger, A.E. Petro, K.L. Macnaul, L.J. Kelly, B.B. Zhang, K. Richards,A. Elbrecht, B.A. Johnson, G. Zhou, T.W. Doebber, C. Biswas, M. Parikh,N. Sharma, M.R. Tanen, G.M. Thompson, J. Ventre, A.D. Adams, R. Mosley,R.S. Surwit, D.E. Moller, Distinct properties and advantages of a novelperoxisome proliferator-activated protein g selective modulator, Mol. Endo-crinol. 17 (2003) 662e676.

[17] M. Goebel, M. Clemenz, B. Staels, T. Unger, U. Kintscher, R. Gust, Character-ization of new PPARg agonists: analysis of Telmisartan's structural compo-nents, ChemMedChem 4 (2009) 445e456.

[18] J.B. Bruning, M.J. Chalmers, S. Prasad, S.A. Busby, T.M. Kamenecka, Y. He,K.W. Nettles, P.R. Griffin, Partial agonists activate PPARgamma using a helix 12independent mechanism, Structure 15 (2007) 1258e1271.

[19] Y. Tanrikulu, B. Krüger, E. Proschak, The holistic integration of virtualscreening in drug discovery, Drug Discov. Today 18 (2013) 358e364.

[20] A. Lavecchia, C. Cerchia, In silico methods to address polypharmacology:current status, applications and future perspectives, Drug Discov. Today 21(2016) 288e298.

[21] F.M.C. da Silva, J.C. dos Santos, J.L.O. Campos, A.C. Mafud, I. Polikarpov,A.C.M. Figueira, A.S. Nascimento, Structure-based identification of novel PPARgamma ligands, Bioorg. Med. Chem. Lett. 23 (2013) 5795e5802.

[22] D. Vidovi�c, S.A. Busby, P.R. Griffin, S.C. Schürer, A combined ligand- andstructure-based virtual screening protocol identifies submicromolar PPARgpartial agonists, ChemMedChem 6 (2011) 94e103.

[23] I.L. Lu, C.-F. Huang, Y.-H. Peng, Y.-T. Lin, H.-P. Hsieh, C.-T. Chen, T.-W. Lien, H.-J. Lee, N. Mahindroo, E. Prakash, A. Yueh, H.-Y. Chen, C.M.V. Goparaju, X. Chen,C.-C. Liao, Y.-S. Chao, J.T.A. Hsu, S.-Y. Wu, Structure-based drug design of anovel family of PPARg partial Agonists: virtual screening, x-ray crystallog-raphy, and in vitro/in vivo biological activities, J. Med. Chem. 49 (2006)2703e2712.

[24] N. Fakhrudin, A. Ladurner, A.G. Atanasov, E.H. Heiss, L. Baumgartner, P. Markt,D. Schuster, E.P. Ellmerer, G. Wolber, J.M. Rollinger, H. Stuppner, V.M. Dirsch,Computer-aided discovery, validation, and mechanistic characterization ofnovel neolignan activators of peroxisome proliferator-activated receptor g,Mol. Pharmacol. 77 (2010) 559e566.

[25] L. Guasch, E. Sala, A. Castell-Auví, L. Ced�o, K.R. Liedl, G. Wolber,M. Muehlbacher, M. Mulero, M. Pinent, A. Ard�evol, C. Valls, G. Pujadas,S. Garcia-Vallv�e, Identification of PPARgamma partial agonists of naturalorigin (I): development of a virtual screening procedure and in vitro valida-tion, PLoS ONE 7 (2012) e50816.

[26] S. Sundriyal, B. Viswanad, P. Ramarao, A.K. Chakraborti, P.V. Bharatam, NewPPARg ligands based on barbituric acid: virtual screening, synthesis and re-ceptor binding studies, Bioorg. Med. Chem. Lett. 18 (2008) 4959e4962.

[27] S. Sundriyal, B. Viswanad, E. Bharathy, P. Ramarao, A.K. Chakraborti,P.V. Bharatam, New PPARg ligands based on 2-Hydroxy-1,4-Naphthoquinone:computer-aided design, synthesis, and receptor-binding studies, Bioorg. Med.Chem. Lett. 18 (2008) 3192e3195.

[28] B.O. Al-Najjar, H.A. Wahab, T.S. Tengku Muhammad, A.C. Shu-Chien,N.A. Ahmad Noruddin, M.O. Taha, Discovery of new nanomolar peroxisomeproliferator-activated receptor g activators via elaborate ligand-basedmodeling, Eur. J. Med. Chem. 46 (2011) 2513e2529.

[29] P.C. Hawkins, A.G. Skillman, A. Nicholls, Comparison of shape-matching anddocking as virtual screening tools, J. Med. Chem. 50 (2007) 74e82.

[30] G. Hu, G. Kuang, W. Xiao, W. Li, G. Liu, Y. Tang, Performance evaluation of 2Dfingerprint and 3D shape similarity methods in virtual screening, J. Chem. Inf.Model. 52 (2012) 1103e1113.

[31] A.S. Karaboga, J.M. Planesas, F. Petronin, J. Teixid�o, M. Souchet, V.I. P�erez-Nueno, Highly specific and sensitive pharmacophore model for identifyingCXCR4 antagonists. Comparison with docking and shape-matching virtualscreening performance, J. Chem. Inf. Model. 53 (2013) 1043e1056.

[32] D.M. Krüger, A. Evers, Comparison of structure- and ligand-based virtualscreening protocols considering hit list complementarity and enrichmentfactors, ChemMedChem 5 (2010) 148e158.

[33] G.B. McGaughey, R.P. Sheridan, C.I. Bayly, J.C. Culberson, C. Kreatsoulas,S. Lindsley, V. Maiorov, J.F. Truchon, W.D. Cornell, Comparison of topological,shape, and docking methods in virtual screening, J. Chem. Inf. Model. 47(2007) 1504e1519.

[34] J. Meslamani, J. Li, J. Sutter, A. Stevens, H.O. Bertrand, D. Rognan, Protein-ligand-based pharmacophores: generation and utility assessment in

computational ligand profiling, J. Chem. Inf. Model. 52 (2012) 943e955.[35] V.I. P�erez-Nueno, D.W. Ritchie, O. Rabal, R. Pascual, J.I. Borrell, J. Teixid�o,

Comparison of ligand-based and receptor-based virtual screening of hiv entryinhibitors for the CXCR4 and CCR5 receptors using 3D ligand shape matchingand Ligand�Receptor docking, J. Chem. Inf. Model. 48 (2008) 509e533.

[36] P. Tiikkainen, P. Markt, G. Wolber, J. Kirchmair, S. Distinto, A. Poso,O. Kallioniemi, Critical comparison of virtual screening methods against theMUV data set, J. Chem. Inf. Model. 49 (2009) 2168e2178.

[37] G. Tresadern, D. Bemporad, T. Howe, A comparison of ligand based virtualscreening methods and application to corticotropin releasing factor 1 recep-tor, J. Mol. Graph. Modell. 27 (2009) 860e870.

[38] V. Venkatraman, V.I. P�erez-Nueno, L. Mavridis, D.W. Ritchie, Comprehensivecomparison of ligand-based virtual screening tools against the DUD data setreveals limitations of current 3D methods, J. Chem. Inf. Model. 50 (2010)2079e2093.

[39] T. Kaserer, V. Temml, Z. Kutil, T. Vanek, P. Landa, D. Schuster, Prospectiveperformance evaluation of selected common virtual screening tools. Casestudy: cyclooxygenase (COX) 1 and 2, Eur. J. Med. Chem. 96 (2015) 445e457.

[40] T. Kaserer, M. H€oferl, K. Müller, S. Elmer, M. Ganzera, W. J€ager, D. Schuster, Insilico predictions of drug-drug interactions caused by CYP1A2, 2C9, and 3A4inhibition - a comparative study of virtual screening performance, Mol. Inf. 34(2015) 431e457.

[41] N.S. Sipes, M.T. Martin, P. Kothiya, D.M. Reif, R.S. Judson, A.M. Richard,K.A. Houck, D.J. Dix, R.J. Kavlock, T.B. Knudsen, Profiling 976 ToxCast chemicalsacross 331 enzymatic and receptor signaling assays, Chem. Res. Toxicol. 26(2013) 878e895.

[42] Dassault Syst�emes, BIOVIA, Pipeline Pilot Release 9.1, Dassault Syst�emes, SanDiego, 2013.

[43] PerkinElmer, ChemBioDraw Ultra 11.0, Waltham, MA, 2008.[44] Accelry Software Inc, Pipeline Pilot Release 8.5, Accelrys Inc, San Diego, 2011.[45] G. Wolber, T. Langer, LigandScout: 3-D pharmacophores derived from

protein-bound ligands and their use as virtual screening filters, J. Chem. Inf.Model. 45 (2005) 160e169.

[46] vROCS Version 3.0.0; OpenEye Scientific Software: Santa FE, NM http://www.eyesopen.com.

[47] G. Jones, P. Willett, R.C. Glen, A.R. Leach, R. Taylor, Development and valida-tion of a genetic algorithm for flexible docking, J. Mol. Biol. 267 (1997)727e748.

[48] GOLD Version 5.2; CCDC: Cambridge, UK. www.ccdc.cam.ac.uk.[49] M.J. Keiser, B.L. Roth, B.N. Armbruster, P. Ernsberger, J.J. Irwin, B.K. Shoichet,

Relating protein pharmacology by ligand chemistry, Nat. Biotechnol. 25(2007) 197e206.

[50] D.A. Filimonov, A.A. Lagunin, T.A. Gloriozova, A.V. Rudik, D.S. Druzhilovskii,P.V. Pogodin, V.V. Poroikov, Prediction of the biological activity spectra oforganic compounds using the pass online web resource, Chem. Heterocycl.Comp. 50 (2014) 444e457.

[51] X. Liu, S. Ouyang, B. Yu, Y. Liu, K. Huang, J. Gong, S. Zheng, Z. Li, H. Li, H. Jiang,PharmMapper server: a web server for potential drug target identificationusing pharmacophore mapping approach, Nucleic Acids Res. 38 (2010)W609eW614.

[52] F.J. Schopfer, M.P. Cole, A.L. Groeger, C.-S. Chen, N.K.H. Khoo, S.R. Woodcock,F. Golin-Bisello, U.N. Motanya, Y. Li, J. Zhang, M.T. Garcia-Barrio, T.K. Rudolph,V. Rudolph, G. Bonacci, P.R.S. Baker, H.E. Xu, C.I. Batthyany, Y.E. Chen,T.M. Hallis, B.A. Freeman, Covalent peroxisome proliferator-activated recep-torg binding by nitro-fatty acids: endogenous ligands act as selective modu-lators, J. Biol. Chem. (2010).

[53] R.C. Braga, C.H. Andrade, Assessing the performance of 3D pharmacophoremodels in virtual screening: how good are they? Curr. Top. Med. Chem. 13(2013) 1127e1138.

[54] N. Triballeau, F. Acher, I. Brabet, J.-P. Pin, H.-O. Bertrand, Virtual screeningworkflow development guided by the “receiver operating characteristic”curve approach. Application to high-throughput docking on metabotropicglutamate receptor subtype 4, J. Med. Chem. 48 (2005) 2534e2547.

[55] E. Lounkine, M.J. Keiser, S. Whitebread, D. Mikhailov, J. Hamon, J.L. Jenkins,P. Lavan, E. Weber, A.K. Doak, S. Cot�e, B.K. Shoichet, L. Urban, Large-scaleprediction and testing of drug activity on side-effect targets, Nature 486(2012) 361e367.

[56] V. Poroikov, D. Akimov, E. Shabelnikova, D. Filimonov, Top 200 medicines: cannew actions be discovered through computer-aided prediction? Sar. QSAREnviron. Res. 12 (2001) 327e344.

[57] M. Jacobsson, P. Lid�en, E. Stjernschantz, H. Bostr€om, U. Norinder, Improvingstructure-based virtual screening by multivariate analysis of scoring data,J. Med. Chem. 46 (2003) 5781e5789.

[58] R.V. Connors, Z. Wang, M. Harrison, A. Zhang, M. Wanska, S. Hiscock, B. Fox,M. Dore, M. Labelle, A. Sudom, S. Johnstone, J. Liu, N.P.C. Walker, A. Chai,K. Siegler, Y. Li, P. Coward, Identification of a PPARd agonist with partialagonistic activity on PPARg, Bioorg. Med. Chem. Lett. 19 (2009) 3550e3554.

[59] S. Lin, Y. Han, Y. Shi, H. Rong, S. Zheng, S. Jin, S.-Y. Lin, S.-C. Lin, Y. Li, Revealinga steroid receptor ligand as a unique PPARg agonist, Cell Res. 22 (2012)746e756.

[60] E. Burgermeister, A. Schnoebelen, A. Flament, J. Benz, M. Stihle, B. Gsell,A. Rufer, A. Ruf, B. Kuhn, H.P. M€arki, J. Mizrahi, E. Sebokova, E. Niesor,M. Meyer, A novel partial agonist of peroxisome proliferator-activated re-ceptor-g (PPARg) recruits PPARg-coactivator-1a, prevents triglyceride accu-mulation, and potentiates insulin signaling in vitro, Mol. Endocrinol. 20 (2006)

T. Kaserer et al. / European Journal of Medicinal Chemistry 124 (2016) 49e6262

809e830.[61] G.Q. Shi, J.F. Dropinski, B.M. McKeever, S. Xu, J.W. Becker, J.P. Berger,

K.L. MacNaul, A. Elbrecht, G. Zhou, T.W. Doebber, P. Wang, Y.-S. Chao,M. Forrest, J.V. Heck, D.E. Moller, A.B. Jones, Design and synthesis of a-ary-loxyphenylacetic acid Derivatives: a novel class of PPARa/g dual agonists withpotent antihyperglycemic and lipid modulating activity, J. Med. Chem. 48(2005) 4457e4468.

[62] C. Weidner, J.C. de Groot, A. Prasad, A. Freiwald, C. Quedenau, M. Kliem,A. Witzke, V. Kodelja, C.-T. Han, S. Giegold, M. Baumann, B. Klebl, K. Siems,L. Müller-Kuhrt, A. Schürmann, R. Schüler, A.F.H. Pfeiffer, F.C. Schroeder,K. Büssow, S. Sauer, Amorfrutins are potent antidiabetic dietary naturalproducts, Proc. Natl. Acad. Sci. U. S. A. 109 (2012) 7257e7262.

[63] A. Furukawa, T. Arita, T. Fukuzaki, S. Satoh, M. Mori, T. Honda, Y. Matsui,K. Wakabayashi, S. Hayashi, K. Araki, J. Ohsumi, Substituents at the naph-thalene C3 position of (�)-Cercosporamide derivatives significantly affect themaximal efficacy as PPARg partial agonists, Bioorg. Med. Chem. Lett. 22 (2012)1348e1351.

[64] J.C. de Groot, C. Weidner, J. Krausze, K. Kawamoto, F.C. Schroeder, S. Sauer,K. Büssow, Structural characterization of amorfrutins bound to the peroxi-some proliferator-activated receptor g, J. Med. Chem. 56 (2013) 1535e1543.

[65] D.B. Kitchen, H. Decornez, J.R. Furr, J. Bajorath, Docking and scoring in virtualscreening for drug discovery: methods and applications, Nat. Rev. Drug Dis-cov. 3 (2004) 935e949.

[66] P.A. Williams, J. Cosme, A. Ward, H.C. Angove, D. Matak Vinkovic, H. Jhoti,Crystal structure of human cytochrome P450 2C9 with bound warfarin, Na-ture 424 (2003) 464e468.

[67] M. Ekroos, T. Sj€ogren, Structural basis for ligand promiscuity in cytochromeP450 3A4, Proc. Natl. Acad. Sci. U. S. A. 103 (2006) 13682e13687.

[68] R.T. Nolte, G.B. Wisely, S. Westin, J.E. Cobb, M.H. Lambert, R. Kurokawa,M.G. Rosenfeld, T.M. Willson, C.K. Glass, M.V. Milburn, Ligand binding and Co-

Activator assembly of the peroxisome proliferator-activated receptor-gamma,Nature 395 (1998) 137e143.

[69] T. Scior, A. Bender, G. Tresadern, J.L. Medina-Franco, K. Martínez-Mayorga,T. Langer, K. Cuanalo-Contreras, D.K. Agrafiotis, Recognizing pitfalls in virtualscreening: a critical review, J. Chem. Inf. Model. 52 (2012) 867e881.

[70] Dassault Syst�emes, BIOVIA, Discovery Studio Modeling Environment Version4.0, Dassault Syst�emes, San Diego, 2013.

[71] V. Temml, T. Kaserer, Z. Kutil, P. Landa, T. Vanek, D. Schuster, Pharmacophoremodelling for Cyclooxygenase-1 and 2 inhibitors with LigandScout in com-parison to discovery Studio, Future Med. Chem. 6 (2014) 1869e1881.

[72] I. Neagu, M. Ohlmeyer, V.M. Paradkar, K.W. Saionz, K. Iwata, T. Okamura,T. Shibutani, Lipoprotein lipase-activating compositions comprising benzenederivatives for treatment of hyperlipidemia, Obes. PCT Int. Appl. (2009). WO2009139076 A1 20091119.

[73] J.C. Medina, D.L. Clark, J.A. Flygare, T.J. Rosen, B. Shan, Preparation of Penta-fluorobenzenesulfonamides for Treating Atherosclerosis and Hypercholes-terolemia, 1999. US 5880151 A 19990309.

[74] D. Clark, J. Flygare, J.C. Medina, T. Rosen, B. Shan, Preparation of penta-fluorobenzenesulfonanilides and analogs as LDL receptor gene expressionregulators, PCT Int. Appl. (1997). WO 9730677 A2 19970828.

[75] J.-L. Chen, Preparation of N-Substituted benzenesulfonamide as PPARg mod-ulators for use in the treatment of neoplasm, osteoporosis, inflammation andrelated diseases, PCT Int. Appl. (2005). WO 2005086904 A2 20050922.

[76] K. Kol�sek, J. Mavri, M. Sollner Dolenc, S. Gobec, S. Turk, Endocrine Dis-ruptomedAn open source prediction tool for assessing endocrine disruptionpotential through nuclear receptor binding, J. Chem. Inf. Model. 54 (2014)1254e1267.

[77] J.J. Mueller, M. Schupp, T. Unger, U. Kintscher, U. Heinemann, Binding Di-versity of Pioglitazone by Peroxisome Proliferator-activated Receptor-gamma.The Protein Data Bank, PDB-code 2XKW, DOI:10.2210/pdb2xkw/pd.