Arch Pharm Res Vol 30, No 5, 533-542, 2007

533

http://apr.psk.or.kr

Pharmacophore Modeling for Protein Tyrosine Phosphatase 1B

Inhibitors

Kavitha Bharatham†, Nagakumar Bharatham†, and Keun Woo Lee

Division of Applied Life Science, Environmental Biotechnology National Core Research Center, Gyeongsang

National University, Jinju 660-701 Korea

(Received September 13, 2006)

A three dimensional chemical feature based pharmacophore model was developed for theinhibitors of protein tyrosine phosphatase 1B (PTP1B) using the CATALYST software, whichwould provide useful knowledge for performing virtual screening to identify new inhibitors tar-geted toward type II diabetes and obesity. A dataset of 27 inhibitors, with diverse structuralproperties, and activities ranging from 0.026 to 600 µM, was selected as a training set. Hypo1,the most reliable quantitative four featured pharmacophore hypothesis, was generated from atraining set composed of compounds with two H-bond acceptors, one hydrophobic aromaticand one ring aromatic features. It has a correlation coefficient, RMSD and cost difference (nullcost-total cost) of 0.946, 0.840 and 65.731, respectively. The best hypothesis (Hypo1) was val-idated using four different methods. Firstly, a cross validation was performed by randomizingthe data using the Cat-Scramble technique. The results confirmed that the pharmacophoremodels generated from the training set were valid. Secondly, a test set of 281 molecules wasscored, with a correlation of 0.882 obtained between the experimental and predicted activities.Hypo1 performed well in correctly discriminating the active and inactive molecules. Thirdly, themodel was investigated by mapping on two PTP1B inhibitors identified by different pharmaceu-tical companies. The Hypo1 model correctly predicted these compounds as being highlyactive. Finally, docking simulations were performed on few compounds to substantiate the roleof the pharmacophore features at the binding site of the protein by analyzing their binding con-formations. These multiple validation approaches provided confidence in the utility of this phar-macophore model as a 3D query for virtual screening to retrieve new chemical entitiesshowing potential as potent PTP1B inhibitors.

Key words: PTP1B Inhibitors, Protein tyrosine phosphatase 1B, Pharmacophore hypothesis,Molecular docking

INTRODUCTION

Phosphorylation and dephosphorylation of proteins actas a switching mechanism to stimulate or disable proteinactivity. Therefore, the balance between protein tyrosinekinases (PTKs) and protein tyrosine phosphatases (PTPs)is very important. Deregulation of a single enzyme couldlead to various diseases. One such example is diabetesmellitus, a common disorder in humans, caused due tothe deregulation of insulin signaling. PTP1B is a negativeregulator of the insulin signaling and thus, would be aspecific target for treating type II diabetes (Johnson et al.,

2002). It has also been proposed as a target in thetreatment of obesity because it regulates leptin signaltransduction (Zabolotny et al., 2002). PTP1B knockoutand antisense studies have shown lower blood glucoselevels and improved insulin responsiveness in normal anddiabetic mice via enhanced insulin receptor signaling inperipheral tissues (Elchebly et al., 1999; Klaman et al.,2000; Zinker et al., 2002). Thus, inhibiting the activity ofPTP1B could be a promising way to treat diseases relatedto its activity. Inhibitors of PTP1B as a target for diabetes,started with vanadium salts, with vanadate found to be apotent, nonselective inhibitor of phosphatases. By themid-1980s, it was understood that blocking one or morephosphatases could enhance the phosphorylation state ofthe insulin receptor kinase/subunit and/or its downstreamsignaling partners and restore the insulin resistance,which is a characteristic of type II diabetes (Hooft van

†Both authors have contributed equally to the work.Correspondence to: Keun Woo Lee, Division of Applied Life Sci-ence, Environmental Biotechnology National Core Research Cen-ter, Gyeongsang National University, Jinju 660-701 KoreaE-mail: kwlee@gnu.ac.kr

534 K. Bharatham et al.

Huijsduijnen et al., 2004). Since then, many drugs havebeen synthesized by various companies for targetingPTP1B, which is very challenging due to the closed formof the catalytic site of PTPs containing a highly polarphosphotyrosine (pTyr) binding site. Presently, compoundswith various activities, drug properties and mechanisms ofaction have been identified. It is interesting to note thatmost of these compounds have diverse modes of action,ranging from the classical competitive type of inhibitors tononcompetitive binders or redox agents. Nevertheless, allcompounds target the active site.

The quest for oral PTP1B inhibitors, with a satisfactorybalance between physicochemical properties and selectivity,is still in its early stages, but despite the recent progress,compounds with optimal oral activity remain to be dis-covered. So far, there has been no report on developingpharmacophore models using inhibitors of PTP1B, whichis currently a powerful tool in identifying new leads. Apharmacophore model represents the 3D arrangementsof the structural or chemical features of a drug (smallorganic compounds, peptides, peptidomimetics, etc.) thatmay be essential for interacting with the protein foroptimum binding. These pharmacophore models can beused differently in drug design programs (i) as a 3D querytool for virtual screening to identify potential new compoundsfrom 3D databases of “drug-like” molecules that havepatentable structures different from those that currentlyexist, and (ii) as a tool to predict the activities of a set ofnew compounds that remain to be synthesized. A largenumber of successful applications have clearly beendemonstrated in medicinal chemistry (Bharatham et al.,2006a). Here, attempts were made to: (i) select a trainingset on a rational basis of the diversities in structures andactivities, (ii) generate a pharmacophore hypothesis basedon the training set, and (iii) validate the pharmacophoremodel using four distinct methods.

METHODS

Biological data collectionAn in-house PTP1B inhibitor database has been col-

lected and developed from various journals using theMDL ISIS software (Shen, 2003). The database is com-prised of 506 PTP1B inhibitors, with both their structureand biological activity. Compounds were eliminated if theylacked exact activity values or had completely differentassays. The 27 compounds (Liljebris et al., 2002;Hamaguchi et al., 2000; Ahn et al., 2002; Chen et al.,2002; Guertin et al., 2003; Lau et al., 2004; Shrestha etal., 2004; Black et al., 2005; Wang et al., 1998; Holmes etal., 2005; Wrobel et al., 1999; Malamas et al., 2000a;Malamas et al., 2000b; Lazo et al., 2001; Hartshorn et al.,2005) were rationally and intuitively selected as a training

set, and represented structural diversity and covered theentire activity range.

Training set selection criteriaThe most critical aspect in the generation of the phar-

macophore hypothesis using the CATALYST software isselection of the training set. Some basic strategies havebeen elegantly laid out, with basic guidelines as follows.(i) A minimum of 16 diverse compounds should beselected to avoid any chance correlation. (ii) The activitydata should have a range of 4-5 orders of magnitude. (iii)The compounds should be selected to provide clear,concise information to avoid redundancy or bias in termsof both structural features and activity range. (iv) Themost active compounds should be included so that theyprovide information on the most critical features requiredfor a reliable/rational pharmacophore model. (v) Inclusionof any compound known to be inactive due to sterichindrance must be avoided because the current featureswithin the CATALYST software cannot handle such cases.On the basis of the above criteria, 27 and 281 compoundswere selected for the training and test sets, respectively.

Pharmacophore hypothesis generationPharmacophore hypotheses generation was achieved

using the HypoGen module of CATALYST 4.10 softwareand the training set of 27 compounds. Molecular flexibilitywas taken into account by considering each compound asa collection of conformers representing a different area ofconformational space accessible to the molecule within agiven energy range. The “best quality searching procedure”was adopted to select representative conformers within a20 kcal/mol range above the computed global minimumenergy. Training set compounds, as shown in Fig. 1, wereimported into CATALYST and conformation modelsgenerated for all molecules (both training and test sets),with 250 maximum number of conformations. CATALYSTselects conformers using the Poling algorithm, whichpenalizes any newly generated conformer if it is too closeto any previously found conformer. This method ensuresmaximum coverage of the conformation space. During theinitial phase of the hypothesis generation exercise, it wasobserved that features, like H-bond acceptors (HBA),hydrophobic aromatic (HY-AR) and ring aromatic (RA),could effectively map all critical chemical/structural featuresof all the training set molecules. Therefore, ten pharma-cophore hypotheses were generated from the training set,using a default uncertainty value of 3. The minimum andmaximum feature counts were 0 and 5, respectively.

Quality of pharmacophore hypothesis assessmentThe quality of the generated pharmacophore hypotheses

was evaluated by considering the cost functions (re-

Pharmacophore Modeling for PTP1B Inhibitors 535

presented in bits unit) calculated using the CATALYST/HypoGen module during hypothesis generation. A mean-ingful pharmacophore hypothesis might be generatedwhen the difference between the null (cost of a complexhypothesis) and the fixed cost (cost of a simple hypothesis)values was large, i.e. 60-70 bits. The total cost (cost of thegenerated hypothesis) should be close to the fixed cost,

and the configuration cost should be less than 17, as thelatter describes the complexity of the hypotheses space tobe explored. Hypo1 satisfied all the above criteria andthus, was chosen for the validation.

Validation of best pharmacophore model (Hypo1)Validation was performed using the following three

Fig. 1. The molecular structures of the 27 training set compounds, taken from the literature

536 K. Bharatham et al.

distinct methods: i) Cat-Scramble program (Du et al.,2005) available in CATALYST, a validation procedurebased on the Fischer’s randomization test, which checkswhether a strong correlation exists between the chemicalstructures and the biological activity. This is carried out byrandomizing the activity data associated with the trainingset compounds. These randomized training sets wereused to generate pharmacophore hypotheses, employingthe same features and parameters as used in thedevelopment of the original pharmacophore hypothesis. Ifthe randomized data set results in the generation of apharmacophore model with similar or better cost values,root mean square deviation (RMSD), and correlationcoefficient (r) of the original hypothesis, then it would havebeen generated by chance. In this validation test, the 95%confidence level was selected and thus, 19 spreadsheetswere generated, ii) a test set of 281 compounds havingexperimental PTP1B inhibitory activities were used toquantify the validity, and iii) candidate molecules werefinally mapped onto the model and their activities pre-dicted.

Molecular dockingThe binding orientations of various compounds were

analyzed using GOLD molecular docking and comparedwith the pharmacophore mapping results as a final valida-tion method. The GOLD 3.0.1 program (Genetic Optimisa-tion for Ligand Docking), from Cambridge CrystallographicData Center, UK, (Jones et al., 1997), uses a geneticalgorithm for docking flexible ligands into protein bindingsites (Bharatham, et al., 2006b). A training set compound,test set and candidate molecule were docked into thePTP1B active site (PDB ID: 1Q6P). A radius of 7.0Åaround the crystal ligand was considered as the activesite. The water of crystallization present within the active

site region was also considered in the docking simulation.The RMSD, annealing parameter of van der Waal’s (vdw)interaction and hydrogen bond interaction were consideredwithin 1.5Å, 4.0Å and 2.5Å, respectively.

RESULTS AND DISCUSSION

Pharmacophore hypothesis generation was achievedusing the training set. Compounds 14 (pyrimido-triazine-piperidine) and 16 (pyridazine) of the training set werepotent candidate molecules, discovered by Roche andBiovitrum, respectively. They were included in the trainingset because their activities were available, which wouldhelp in generating a reasonable pharmacophore model.Ten pharmacophore hypotheses were generated, amongwhich the best model was selected. All ten hypotheseshad the same four chemical features. Therefore, selectionof the ideal pharmacophore hypothesis was characterizedusing the high cost difference (null-fixed), low error cost,low RMSD and best correlation coefficient. Hypo1 isconsidered the best as it is characterized by the highestcost difference (72.026), lowest RMSD value (0.846) andalso had the best correlation coefficient (0.946), whichindicates a true correlation and good predictive capability.The total cost value of each hypothesis was close to thefixed cost value, which is expected for a good hypothesis.The configuration cost value of the hypothesis was alsowithin the allowed range, i.e. 17. The null cost, fixed costand the configuration cost values for the 10 best rankinghypotheses were 182.256, 106.97 and 15.028, respectively.The cost values, correlation coefficients (r) for the trainingset, RMSD values and features of all ten pharmacophoremodels are listed in Table I.

Hypo1 comprised of two H-bond acceptors (HBA), onehydrophobic aromatic (HY-AR) and one ring aromatic

Table I. Information of the cost values measured in bits RMSD, correlation coefficient values and features for top-ten hypothesesa

Hypo-thesis Total costCost difference

(null-total cost)RMSD r Features

Correlation value for

281 test set compounds

1 116.525 65.731 0.840 0.946 HBA,HBA,HY-AR,RA 0.882

2 117.323 64.933 0.874 0.941 HBA,HBA,HY-AR,RA 0.862

3 121.179 61.077 1.025 0.919 HBA,HBA,HY-AR,RA 0.829

4 121.247 61.009 1.028 0.918 HBA,HBA,HY-AR,RA 0.838

5 121.251 61.005 1.021 0.919 HBA,HBA,HY-AR,RA 0.798

6 121.376 60.880 1.024 0.919 HBA,HBA,HY-AR,RA 0.870

7 122.020 60.236 1.054 0.914 HBA,HBA,HY-AR,RA 0.807

8 122.964 59.292 1.087 0.908 HBA,HBA,HY-AR,RA 0.834

9 123.025 59.231 1.083 0.909 HBA,HBA,HY-AR,RA 0.743

10 123.195 59.061 1.094 0.907 HBA,HBA,HY-AR,RA 0.835

aNull cost of top-ten score hypotheses is 182.256 bits Fixed cost is 106.97 bits. Configuration cost is 15.028 bits. Abbreviation used for features:

HBA, hydrogen-bond acceptor; HY-AR, hydrophobic aromatic; RA, ring aromatic.

Pharmacophore Modeling for PTP1B Inhibitors 537

(RA) features (Fig. 2). The two best active training setcompounds that were mapped on to the pharmacophoremodel, Hypo1, are shown in Fig. 4a and 4b, respectively.Compound 2 was a reversible, competitive and catalyticsite-binding PTP1B inhibitor, where the phenyldifluoro-methylphosphonate group targets the primary binding siteand forms a hydrogen bond with Arg221 (Lau et al., 2004,Giovanna Scapin et al., 2003). Analogously, our pharma-cophore model mapped the phosphonate hydroxyls ofcompound 2 as two HBA features, which interact withcatalytic site residues. The docking results for compound2 also revealed that it interacts with Arg221 and Gly220

via hydrogen bonding (Fig. 6a). The hydrophobic phenyl(mapped as RA feature) and indole groups (mapped asHY-AR feature) interact with the Tyr46 and Arg47 sidechains, respectively, which are the secondary binding sites,which have also been extensively studied. Compound 3belongs to a class of highly hydrophobic, more promiscuousPTP1B inhibitors, with a benzothiophene biphenyl oxo-acetic acid group. The carboxylic acid portion of the liganddoes not interact directly with Arg221, but participates inhydrogen bonds with two water molecules bridging theligand with Arg221 (Malamas et al., 2000b). The two HBAfeatures of our pharmacophore model were mapped ontothe carboxyl group, which indirectly interacts with catalyticsite residues.

A pharmacophore model capable of predicting an activeor inactive compound would reduce the time of the drugdiscovery process. Therefore, all compounds in both thetraining and test sets were classified into three groups;highly active (+++, IC50 1 µM), moderately active (++, 1µM < IC50 < 10 µM) and inactive compounds (+, 10 µM).The experimental and Hypo1 predicted activities for the27 training set compounds are shown in Table II. Out ofthe 27 compounds, only two moderately active (++)compounds were predicted as being inactive (+), while allthe remaining compounds were predicted to the samemagnitude. The error value was calculated as the ratiobetween the predicted and experimental activities. Apositive error value indicates that the predicted IC50 ishigher than that obtained experimentally, while a negativeerror value indicates that the predicted IC50 is lower thanthat obtained experimentally. An error value of less thanten represents a difference no greater than one orderbetween the predicted and experimental activities. Amongthe training set compounds, very few had an error value >3. Hence, the pharmacophore model was selected for

Fig. 2. The best ranked pharmacophore hypothesis (Hypo1). The distances between the chemical features of the 3D pharmacophore hypothesis,

where orange represents the ring aromatic (RA), green the H-bond acceptor (HBA) and light blue the hydrophobic aromatic (HY-AR) feature.

Fig. 3. Correlation graph between the experimental and Hypo1

predicted activities

538 K. Bharatham et al.

validation using various methods.The validation of Hypo1 was performed using four distinct

methods. First, a validation procedure was performedusing the Cat-Scramble technique, which is based on theprinciple of randomizing the activity data of the training setcompounds, and generates pharmacophore hypothesesusing the same features and parameters as used in thedevelopment of the original pharmacophore hypothesis.Nineteen random spreadsheets (or 19 HypoGen runs)were generated for the 95% confidence level. The data ofcross validation clearly indicated that all values generatedafter randomization produced hypotheses with no pre-dictive value near or similar to that obtained for Hypo1, asshown in Table III. Out of the 19 runs, only two had acorrelation close to 0.7, but the RMSD was high and thetotal cost was close to the null cost, which is not desirablefor a good hypothesis. This cross validation providedstrong confidence on the initial pharmacophore model,Hypo1.

The validity of any pharmacophore model needs to beascertained by its application to the test set to find if the

model correctly predicts the activity of the molecules inthe test set molecules and most importantly, whether itcorrectly identifies active and inactive molecules. Therefore,in the second method, 281 molecules with experimentalPTP1B inhibitory activities were imported, and conformersgenerated in a similar manner to that of the training set. Aregression analysis was performed on the 281 test setcompounds, with the best hypothesis (Hypo1) generatedfrom the original training set. The graph drawn betweenexperimental and predicted activities generated for thetest set compounds using Hypo1 is shown in Fig. 3. Acorrelation of 0.882 was observed between the experi-mental and predicted activity values, signifying a goodcorrelation. None of the test set compounds had an errorvalue greater than 10, implying their activities were predi-cted within the same magnitude.

Two of the highly active test set compounds weremapped onto Hypo1 (Fig. 4c and 4d). The molecularstructures of the two test set compounds were shown inFig. 5a and 5b. The predicted activities of t188 and t362were 0.48 µM (experimental IC50 = 1.01 µM), and 0.15 µM

Fig. 4. Mapping of the top scored Hypo1 on the highly active training set compounds 2 (a) and 3 (b), the test set compounds t188 (c) and t362 (d),

and the candidate molecules c1 (e) and c2 (f). In the pharmacophore hypothesis; orange represents the ring aromatic (RA), green the H-bond

acceptor (HBA) and light blue the hydrophobic aromatic (HY-AR) feature.

Pharmacophore Modeling for PTP1B Inhibitors 539

(experimental IC50 = 0.145 µM), respectively. Although nospecific binding mechanism has been described for 1, 2-naphthoquinone derivatives, it was assumed that the acidgroup of t188 forms H-bonding with Arg221 (Ahn et al.,2002). Accordingly, the two HBA features of our pharma-cophore model were mapped perfectly onto the acidgroup of t188. The docking results also revealed that thecarboxylic group formed hydrogen bonds with the guanidogroup of Arg221. The RA feature, when mapped onto 1,2-naphthoquinone, forms hydrophobic interactions with

the side chain of Arg47, while the HY-AR feature whenmapped onto the phenyl ring of the indole group formshydrophobic interactions with Met258. The binding orien-tation of t188 at the active site is shown in Fig. 6b. In

Table II. Experimental and predicted biological data measured asIC50 (µM) based on pharmacophore model Hypo1 for training setmolecules

Comp.Experimental

Activity (µM)

Pre.

Activity Error a Fitness

score b

Activity

Scale c

Pre. Activity

Scale

1 0.026 0.038 1.5 6.93 +++ +++

2 0.038 0.085 2.2 6.58 +++ +++

3 0.052 0.012 –4.2 7.41 +++ +++

4 0.061 0.242 3.9 6.13 +++ +++

5 0.098 0.079 –1.2 6.61 +++ +++

6 0.16 0.29 1.8 6.05 +++ +++

7 0.24 0.16 –1.5 6.31 +++ +++

8 0.39 0.55 1.4 5.76 +++ +++

9 0.65 0.76 1.2 5.63 +++ +++

10 0.85 0.59 –1.4 5.74 +++ +++

11 1.3 1.1 –1.2 5.46 ++ ++

12 1.9 1.6 –1.2 5.29 ++ ++

13 2.1 1 –2.1 5.51 ++ ++

14 2.9 2.3 –1.2 5.14 ++ ++

15 4.3 8.9 2.1 4.56 ++ ++

16 5.6 82 15 3.60 ++ +

17 6 5.2 –1.2 4.8 ++ ++

18 8 38 4.8 3.92 ++ +

19 26 47 1.8 3.84 + +

20 42 40 –1 3.9 + +

21 55 38 –1.5 3.93 + +

22 73 60 –1.2 3.73 + +

23 86 58 –1.5 3.75 + +

24 95 58 –1.6 3.74 + +

25 120 39 –3 3.92 + +

26 140 63 –2.2 3.71 + +

27 600 83 –7.2 3.59 + +

a + indicates that the predicted IC50 is higher than the experimental

IC50; – indicates that the predicted IC50 is lower than the experimental

IC50

b Fitness score indicates how well the features in the pharmacophore

overlap the chemical features in the molecule.C PTP1B activity scale: +++, IC50 1 µM (high active); ++, 1 µM < IC50 <

10 µM (moderate active); +, IC50 10 µM (inactive)

Table III. Results from cross-validation using CATSCRAMBLEmethod in CATALYST a

Trial no. Total cost Fixed cost RMSD Correlation

coefficient (r)

Results for unscrambled

116.525 106.97 0.840 0.946

Results for scrambled

1 169.943 100.237 2.260 0.496

2 171.293 101.267 2.269 0.490

3 182.256 90.817 2.600 *

4 166.227 104.404 2.133 0.573

5 164.798 101.142 2.171 0.551

6 179.061 98.481 2.411 0.380

7 172.685 103.646 2.258 0.497

8 178.817 96.942 2.445 0.345

9 165.570 103.446 2.145 0.566

10 182.256 90.813 2.602 *

11 175.474 101.531 2.329 0.447

12 164.358 106.077 2.076 0.602

13 177.817 103.612 2.325 0.451

14 182.256 90.817 2.602 *

15 165.038 101.921 2.162 0.556

16 182.256 90.817 2.602 *

17 149.642 107.991 1.756 0.737

18 178.492 99.557 2.412 0.376

19 147.243 103.432 1.801 0.721

aNull cost is 182.256

*Indicates that correlation coefficient could not be obtained.

Fig. 5. The molecular structures of test set compounds t188 (a) and

t362 (b) and the candidate molecules c1 (c) and c2 (d)

540 K. Bharatham et al.

compound t362, the phosphonate group was earlierhypothesized to bind to the side chain of the active siteArg221 via a charge-charge interaction (Wrobel et al.,1999). The docking results also confirm the hydroxylgroups of phosphonate form hydrogen bonds with Arg221(guanido group) and Gly220 (main chain). The HY-ARfeature (mapped on corner phenyl of fused ring) formedhydrophobic interactions with the Arg47 side chain, whilethe RA feature mapped onto the five membered ringformed hydrophobic interactions with Tyr46 and Lys120(Fig. 6c). As the pharmacophore model represented thecrucial features required for binding, it was able toappropriately predict their activities.

As a third trial, another validation process was performedto find the usefulness of the selected pharmacophoremodel in evaluating two diverse potent PTP1B inhibitors(Fig. 5c and 5d), which have either been clinically intro-duced or are in the developmental stages. The rationaleof this approach was if the pharmacophore model mapsonto those inhibitors and predicts their activity well, thepharmacophore model is then expected to be useful as asearch tool for identifying new PTP1B inhibitors. No

attempts were made to directly compare the predictedand experimental activities of these compounds, as nodetailed activity data have previously been reported. MerckFrosst has described a charged compound, monodifluoro-phosphonate derivative (c1), but no biological data wasdisclosed. Hypo1 predicted the activity (IC50 for binding toPTP1B) to be 0.0061 µM, and also correctly classified themolecule as being highly active. The two HBAs map theoxygen atoms of the phosphonate involved in the catalyticactivity. The RA and HY-AR features map the phenyl ringsthat contribute to the hydrophobic interactions. Similarly,the docking results also showed the two hydroxyl groupsof phosphonate formed hydrogen bonds with Arg221 andGly220, while the benzoyl and phenyl groups formedhydrophobic interactions with the Arg47 side chain andMet258, respectively (Fig. 6d).

To avoid multiple negative charges, while still showingpotency, several inhibitors, such as ertiprotafib (c2), showedmodest selectivity profiles, have been tested in clinicaltrials and have reached phase II development. Theinhibition of PTP1B by this compound was presumed toinvolve the oxidation of the active site cysteine of PTP1B

Fig. 6. The molecular docking results. Docked compound 2 of the training set (a), the test set compounds t188 (b) and t362 (c), and the candidate

molecule c1 (d) with the catalytic residues are represented as sticks, with water as a sphere in the PTP1B protein active site.

Pharmacophore Modeling for PTP1B Inhibitors 541

to the corresponding sulfenic acid. The Hypo1 predictedactivity was 0.072 µM, and the molecule was correctlyclassified as being highly active. The two HBAs map theoxygen atoms of the carboxylate involved in the catalyticactivity. The RA and HY-AR features map the fusedthiophene and phenyl rings of 11-Arylbenzo[b]naphtho[2,3-d]thiophenes, respectively, which contribute to thehydrophobic interactions. The mapping of the candidatemolecules, c1 and c2, using Hypo1 are shown in Fig. 4eand 4f. As a fourth validation method, docking studieswere performed, and the results were simultaneouslydescribed for better comparison with the pharmacophoremodel mapping. Thus, Hypo1 was validated using fourmethods, all of which gave reliable results. Therefore, thepharmacophore model is expected to help in the identi-fication of new classes of PTP1B inhibitors from virtualscreening.

CONCLUSIONS

One of the major goals of this study was to generate apredictive pharmacophore model that could be utilized asa query tool to search 3D databases of diverse drug-likecompounds for the identification of new molecules thatpossess potent PTP1B inhibitory activities. The best phar-macophore hypothesis, Hypo1, as characterized by thehigh correlation coefficient of 0.946, was generated forPTP1B inhibitors using a training set of 27 compounds,and validated using four distinct methods. In addition, thisstudy clearly indicates that the selected pharmacophoremodel can be used to find new chemical entities withpotent activities against a target disease, which is ofutmost importance to the pharmaceutical industry. Ourpharmacophore model successfully predicted the bio-logical activities of the compounds in the test set, as wellas accurately classified two existing PTP1B inhibitors fromtwo different pharmaceutical companies. The bindingorientations of the compounds also substantiated thefeatures mapped by the pharmacophore model. It hasbeen confirmed that correct mapping to four features issufficient to successfully identify specific PTP1B inhibitors.The selected pharmacophore model is expected to helpidentify new classes of PTP1B inhibitor.

ACKNOWLEDGEMENTS

Nagakumar Bharatham and Kavitha Bharatham wererecipients of fellowships from the BK21 Program. Thiswork was supported by grants from the MOST/KOSEF forthe Environmental Biotechnology National Core ResearchCenter (grant #:R15-2003-012-02001-0) and for the BasicResearch Program (grant #: R01-2005-000-10373-0).

REFERENCES

Ahn, J. H., Cho, S. Y., Ha, J. D., Chu, S. Y., Jung, S. H., Jung, Y.

S., Baek, J. K.,Ki Choi, I., Shin, E. Y., Kang, S. K., Kim, S. S.,

Cheon, H. G., Yang, S. D., and Choi, J. K., Synthesis and

PTP1B inhibition of 1,2-naphthoquinone derivatives as

potent anti-diabetic agents. Bioorg. Med. Chem. Lett., 12,

1941-1946 (2002).

Bharatham, N., Bharatham, K., and Lee, K. W., Pharmacophore

identification and virtual screening for methionyl-tRNA-

synthetase inhibitors. J. Mol. Graph. Model., doi:10.1016/

j.jmgm.2006.08.002.

Bharatham, N., Bharatham, K., and Lee, K. W., Quantitative

Structure-Activity Relationships (QSAR) and molecular

docking studies of P56 LCK inhibitors. Bull. Korean Chem.

Soc., 27, 266-272 (2006).

Black, E., Breed, J., Breeze, A. L., Embrey, K., Garcia, R., Gero,

T. W., Godfrey, L., Kenny, P. W., Morley, A. D., Minshull, C.

A., Pannifer, A. D., Read, J., Rees, A., Russell, D. J., Toader,

D., and Tucker, J., Structure-based design of protein tyrosine

phosphatase-1B inhibitors. Bioorg. Med. Chem. Lett., 15,

2503-2507 (2005).

Chen, R. M., Hu, L. H., An, T. Y., Li, J., and Shen, Q., Natural

PTP1B inhibitors from Broussonetia papyrifera. Bioorg. Med.

Chem. Lett., 12, 3387-3390 (2002).

Combs, A. P., Yue, E. W., Bower, M., Ala, P. J., Wayland, B.,

Douty, B., Takvorian, A., Polam, P., Wasserman, Z., Zhu, W.,

Crawley, M. L., Pruitt, J., Sparks, R Glass, B Modi, D

McLaughlin, E., Bostrom, L., Li, M., Galya, L., Blom, K.,

Hillman, M., Gonneville, L., Reid, B. G., Wei, M., Becker-

Pasha, M., Klabe, R., Huber, R., Li, Y., Hollis, G., Burn, T. C.,

Wynn, R., Liu, P., and Metcalf, B., Structure-based design

and discovery of protein tyrosine phosphatase inhibitors

incorporating novel isothiazolidinone heterocyclic phospho-

tyrosine mimetics. J. Med. Chem., 48, 6544-6548 (2005).

Du, L. P., Li, M. Y., Tsai, K. C., You, Q. D., and Xia, L.,

Characterization of binding site of closed-state KCNQ1

potassium channel by homology modeling, molecular

docking, and pharmacophore identification. Biochem.

Biophys. Res. Commun., 332, 677-687 (2005).

Elchebly, M., Payette, P., Michaliszyn, E., Cromlish, W., Collins,

S., Loy, A. L., Normandin, D., Cheng, A., Himms-Hagen, J.,

Chan, C. C., Ramachandran, C., Gresser, M. J., Tremblay,

M. L., and Kennedy, B. P., Increased insulin sensitivity and

obesity resistance in mice lacking the protein tyrosine

phosphatase-1B gene. Science, 283, 1544-15448 (1999).

Giovanna Scapin, Sangita Patel, B., Joseph Becker, W.,

Qingping Wang, Caroline Desponts, Deena Waddleton,

Kathryn Skorey, Wanda Cromlish, Christopher Bayly, Michel

Therien, Jacques Yves Gauthier, Chun Sing Li, Cheuk Lau,

K., Chidambaram Ramachandran, Brian Kennedy, P., and

Ernest Asante-Appiah, The Structural Basis for the Selec-

tivity of Benzotriazole Inhibitors of PTP1B. Biochemistry, 42,

542 K. Bharatham et al.

11451-11459 (2003).

Guertin, K. R., Setti, L., Qi, L., Dunsdon, R. M., Dymock, B. W.,

Jones, P. S., Overton, H., Taylor, M., Williams, G., Sergi, J. A.,

Wang, K., Peng, Y., Renzetti, M., Boyce, R., Falcioni, F.,

Garippa, R., and Olivier, A. R., Identification of a novel class of

orally active pyrimido[5,4-3][1,2,4]triazine-5,7-diamine-based

hypoglycemic agents with protein tyrosine phosphatase inhi-

bitory activity. Bioorg. Med. Chem. Lett., 13, 2895-2898 (2003).

Hamaguchi, T., Takahashi, A., Kagamizono, T., Manaka, A.,

Sato, M., and Osada, H., Synthesis and characterization of a

potent and selective protein tyrosine phosphatase inhibitor.

Bioorg. Med. Chem. Lett., 10, 2657-2660 (2000).

Hartshorn, M. J., Murray, C. W., Cleasby, A., Frederickson, M.,

Tickle, I. J., and Jhoti, H., Fragment-based lead discovery using

X-ray crystallography. J. Med. Chem., 48, 403-413 (2005).

Holmes, C. P., Li, X., Pan, Y., Xu, C., Bhandari, A., Moody, C.

M., Miguel, J. A., Ferla, S. W., De Francisco, M. N.,

Frederick, B. T., Zhou, S., Macher, N., Jang, L., Irvine, J. D.,

and Grove, J. R., Discovery and structure-activity relationships

of novel sulfonamides as potent PTP1B inhibitors. Bioorg.

Med. Chem. Lett., 15, 4336-4341 (2005).

Hooft van Huijsduijnen, R., Sauer, W. H., Bombrun, A., and

Swinnen, D., Prospects for inhibitors of protein tyrosine

phosphatase 1B as antidiabetic drugs. J. Med. Chem., 47,

4142-4146 (2004).

Johnson, T. O., Ermolieff, J., and Jirousek, M. R., Protein

tyrosine phosphatase 1B inhibitors for diabetes. Nat. Rev.

Drug Discov., 1, 696-709 (2002).

Jones, G., Willett, P., Glen, R. C., Leach, A. R., and Taylor, R.,

Development and validation of a genetic algorithm for flexible

docking. J. Mol. Biol., 267, 727-748 (1997).

Klaman, L. D., Boss, O., Peroni, O. D., Kim, J. K., Martino, J. L.,

Zabolotny, J. M., Moghal, N., Lubkin, M., Kim, Y. B., Sharpe,

A. H., Stricker-Krongrad, A., Shulman, G. I., Neel, B. G., and

Kahn, B. B., Increased energy expenditure, decreased

adiposity, and tissue-specific insulin sensitivity in protein-

tyrosine phosphatase 1B-deficient mice. Mol. Cell Biol., 20,

5479-5489 (2000).

Lau, C. K., Bayly, C. I., Gauthier, J. Y., Li, C. S., Therien, M.,

Asante-Appiah, E., Cromlish, W., Boie, Y., Forghani, F.,

Desmarais, S., Wang, Q., Skorey, K., Waddleton, D.,

Payette, P., Ramachandran, C., Kennedy, B. P., and Scapin,

G., Structure based design of a series of potent and selective

non peptidic PTP-1B inhibitors. Bioorg. Med. Chem. Lett., 14,

1043-1048 (2004).

Lazo, J. S., Aslan, D. C., Southwick, E. C., Cooley, K. A.,

Ducruet, A. P., Joo, B., Vogt, A., and Wipf, P., Discovery and

biological evaluation of a new family of potent inhibitors of the

dual specificity protein phosphatase Cdc25. J. Med. Chem.,

44, 4042-4049 (2001).

Liljebris, C., Martinsson, J., Tedenborg, L., Williams, M., Barker,

E., Duffy, J. E., Nygren, A., and James, S., Synthesis and

biological activity of a novel class of pyridazine analogues as

non-competitive reversible inhibitors of protein tyrosine

phosphatase 1B (PTP1B). Bioorg. Med. Chem., 10, 3197-

3212 (2002).

Malamas, M. S., Sredy, J., Gunawan, I., Mihan, B., Sawicki, D.

R., Seestaller, L., Sullivan, D., and Flam, B. R., New

azolidinediones as inhibitors of protein tyrosine phosphatase

1B with antihyperglycemic properties., J. Med. Chem., 43,

995-1010 (2000).

Malamas, M. S., Sredy, J., Moxham, C., Katz, A., Xu, W.,

McDevitt, R., Adebayo, F. O., Sawicki, D. R., Seestaller, L.,

Sullivan, D., and Taylor, J. R., Novel benzofuran and

benzothiophene biphenyls as inhibitors of protein tyrosine

phosphatase 1B with antihyperglycemic properties. J. Med.

Chem., 43, 1293-1310 (2000).

Moller, N. P. H., Iversen, L. F., Andersen, H. S., and McCormack,

J. M., Protein tyrosine phosphatases (PTPs) as drug targets:

Inhibitors of PTP-1B for the treatment of diabetes. Curr. Opin.

Drug DiscoVery DeV., 3, 527-540 (2000).

Shen, J., HAD: an automated database tool for analyzing

screening hits in drug discovery. J. Chem. Inf. Comput. Sci.,

43, 1668-1672 (2003).

Shrestha, S., Shim, Y. S., Kim, K. C., Lee, K. H., and Cho, H.,

Evans Blue and other dyes as protein tyrosine phosphatase

inhibitors. Bioorg. Med. Chem. Lett., 14, 1923-1926 (2004).

Tonks, N. K. and Neel, B. G., Combinatorial control of the

specificity of protein tyrosine phosphatases. Curr. Opin. Cell

Biol., 13, 182-195 (2001).

Wang, Q., Huang, Z., Ramachandran, C., Dinaut, A. N., and

Taylor, S. D., Naphthalenebis[alpha,alpha-difluoromethy-

lenephosphonates] as potent inhibitors of protein tyrosine

phosphatases. Bioorg. Med. Chem. Lett., 8, 345-350 (1998).

Wrobel, J., Sredy, J., Moxham, C., Dietrich, A., Li, J., Sawicki,

D. R., Seestaller, L., Wu, L., Katz, A., Sullivan, D., Tio, C.,

and Zhang, Z. Y., PTP1B inhibition and antihyperglycemic

activity in the ob/ob mouse model of novel 11-arylbenzo

[b]naphtho[2,3-d]furans and 11-arylbenzo[b]naphtho[2,3-

d]thiophenes. J. Med. Chem., 42, 3199-3202 (1999).

Zabolotny, J. M., Bence-Hanulec, K. K., Stricker-Krongrad, A.,

Haj, F., Wang, Y., Minokoshi, Y., Kim, Y. B., Elmquist, J. K.,

Tartaglia, L. A., Kahn, B. B., and Neel, B. G., PTP1B regulates

leptin signal transduction in vivo. Dev. Cell, 2, 489-495 (2002).

Zhang, Z. Y., Protein tyrosine phosphatases: structure and

function, substrate specificity, and inhibitor development.

Annu. Rev. Pharmacol. Toxicol., 42, 209-234 (2002).

Zinker, B. A., Rondinone, C. M., Trevillyan, J. M., Gum, R. J.,

Clampit, J. E., Waring, J. F., Xie, N., Wilcox, D., Jacobson, P.,

Frost, L., Kroeger, P. E., Reilly, R. M., Koterski, S.,

Opgenorth, T. J., Ulrich, R. J., Crosby, S., Butler, M., Murray,

S. F., McKay, R. A., Bhanot, S., Monia, B. P., and Jirousek,

M. R., PTP1B antisense oligonucleotide lowers PTP1B

protein, normalizes blood glucose, and improves insulin

sensitivity in diabetic mice. Proc. Natl. Acad. Sci. U.S.A., 99,

11357-11362 (2002).