Discovery of new nanomolar peroxisome proliferator-activated receptor γ activators via elaborate ligand-based modeling.

Abstract

Peroxisome Proliferator-Activated Receptor γ (PPARγ) activators have drawn great recent attention in the clinical management of type 2 diabetes mellitus, prompting several attempts to discover and optimize new PPARγ activators. With this in mind, we explored the pharmacophoric space of PPARγ using seven diverse sets of activators. Subsequently, genetic algorithm and multiple linear regression analysis were employed to select an optimal combination of pharmacophoric models and 2D physicochemical descriptors capable of accessing self-consistent and predictive quantitative structure-activity relationship (QSAR) (r2(71)=0.80, F=270.3, r2LOO=0.73, r2PRESS against 17 external test inhibitors=0.67). Three orthogonal pharmacophores emerged in the QSAR equation and were validated by receiver operating characteristic (ROC) curves analysis. The models were then used to screen the national cancer institute (NCI) list of compounds. The highest-ranking hits were tested in vitro. The most potent hits illustrated EC50 values of 15 and 224 nM.

Full text

Page 1

Original article
Discovery of new nanomolar peroxisome proliferator-activated
receptor g activators via elaborate ligand-based modeling
Belal O. Al-Najjara, Habibah A. Wahaba,b,**, Tengku Sifzizul Tengku Muhammadb,c,
Alexander Chong Shu-Chienb,d, Nur Adelina Ahmad Noruddinb, Mutasem O. Tahae,*
aPharmaceutical Design and Simulation (PhDS) Laboratory, School of Pharmaceutical Sciences, Universiti Sains Malaysia, 11800 Minden, Pulau Pinang, Malaysia
bMalaysian Institute of Pharmaceuticals and Nutraceuticals, Ministry of Science, Technology and Innovation, SAINS@USM, No. 10, 11900 Persiaran Bukit Jambul, Pulau Pinang,
Malaysia
cDepartment of Biological Sciences, Universiti Malaysia Terengganu, 21030 Kuala Terengganu, Terengganu, Malaysia
dSchool of Biological Sciences, Universiti Sains Malaysia, 11800, Penang, Malaysia
eDrug Discovery Unit, Department of Pharmaceutical Sciences, Faculty of Pharmacy, University of Jordan, Queen Rania Street, Amman, Jordan
a r t i c l e i n f o
Article history:
Received 4 January 2011
Received in revised form
12 March 2011
Accepted 16 March 2011
Available online xxx
Keywords:
Peroxisome Proliferator-activated receptor g
Pharmacophore modeling
Quantitative structureeactivity relationship
In silico screening
Type 2 diabetes mellitus
a b s t r a c t
Peroxisome Proliferator-Activated Receptor g (PPARg) activators have drawn great recent attention in the
clinical management of type 2 diabetes mellitus, prompting several attempts to discover and optimize
new PPARg activators. With this in mind, we explored the pharmacophoric space of PPARg using seven
diverse sets of activators. Subsequently, genetic algorithm and multiple linear regression analysis were
employed to select an optimal combination of pharmacophoric models and 2D physicochemical
descriptors capable of accessing self-consistent and predictive quantitative structureeactivity relation-
ship (QSAR) (r271¼0.80, F ¼270.3, r2LOO¼0.73, r2PRESSagainst 17 external test inhibitors ¼0.67). Three
orthogonal pharmacophores emerged in the QSAR equation and were validated by receiver operating
characteristic (ROC) curves analysis. The models were then used to screen the national cancer institute
(NCI) list of compounds. The highest-ranking hits were tested in vitro. The most potent hits illustrated
EC50values of 15 and 224 nM.
? 2011 Elsevier Masson SAS. All rights reserved.
1. Introduction
Type II diabetes is a progressive disease characterized by insulin
resistance in peripheral tissues and/or impaired insulin secretion
by the pancreas. The resultant high blood glucose level generally
leads to several serious complications. The World Health Organi-
zation recently warned that type II diabetes has become global
pandemic [1]. Type II diabetes is strongly associated with obesity.
The common link between obesity and type II diabetes is insulin
resistance [1,2]. At the molecular level, the mechanism of insulin
resistance in type II diabetes appears to involve defects in post-
receptor signal transduction [3e5].
PPARs are type II nuclear hormone receptors that participate in
the regulation of fattyacids, carbohydrates and glucose metabolism
[6]. PPARs are ligand-dependent transcriptional regulators heter-
odimerized with retinoid X receptor. They bind to peroxisome
proliferator-response element (PPRE) [7] and regulate the activities
of nuclear factors essential in immunomodulatory and inflamma-
tory responses [8,9]. Three PPAR isoforms have been identified: a,
b (or d), and g [10].
PPARg agonists have drawn great attention in the clinical
management of cardiovascular risk factors associated with meta-
bolic syndrome and type 2 diabetes [11]. In the nineties, several
classes of PPARg agonists were reported to have anti-diabetic
actions, including: thiazolidindiones, dihydrobenzofurans, dihy-
drobenzopyrans, benzofurans benzoxazoles and a-amino-b-phe-
nylpropanoic acid derivatives [12]. Recently, more diverse PPARg
agonists were developed as potential hypoglycemic agents,
including: b-carboxyethyl-rhodanines [13], indene N-oxides [14],
indoles [15], 7-azaindoles [16], 2-aryl-4-oxazolylmethoxy-benzyl-
glycinesand2-aryl-4-thiazolyl-methoxy-benzylglycines
5-arylthiazolidine-2,4-diones [18], aryloxazolidinediones [19] and
phenylacetic acid derivatives [20].
Clearly, the main focus of recent efforts towards the develop-
ment of new PPARg agonists concentrate on structure-based ligand
design [13,21e33] with few ligand-based exceptions [34e37]. To
date, many PPARg X-ray complexes are documented in the Protein
Data Bank [38e47].
[17],
* Corresponding author. Tel.: þ962 6 5355000x23305; fax: þ962 6 5339649.
** Corresponding author. Tel.: þ60164410188.
E-mail addresses: habibahwahab@yahoo.co.uk (H.A. Wahab), mutasem@ju.edu.
jo (M.O. Taha).
Contents lists available at ScienceDirect
European Journal of Medicinal Chemistry
journal homepage: http://www.elsevier.com/locate/ejmech
0223-5234/$ e see front matter ? 2011 Elsevier Masson SAS. All rights reserved.
doi:10.1016/j.ejmech.2011.03.040
European Journal of Medicinal Chemistry xxx (2011) 1e17
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 2

X
N
Y
O
R1
R3
R2
Z
N
R3
Y
O
R1
R2
X
1-2223-32
SO
O
OH
R1
OH
R2
O
O
R3
X
OH
R2
R1
O
R4
33-3536-43
OR1
R2
CH2(n)
O
NH
Y
O
O
SO
X
Y
R1
OHO
Cl
44-7677-80
O
O
O
F
F
F
O
OH
O
O
O
OH
Cl
O
81 82
S
N
H
O
O
O
O
O
S
N
O
O
O
N
N
CH3
8385
S
N
H
O
O
O
O
O
F
S
N
H
O
O
O
N
N
84 86
O
O
O
O
Cl
F
O
OH
N
O
O
O
OH
O
O
F
F
F
8788
Fig. 1. The chemical scaffolds of training compounds, the detailed structures are as in Table A under Supplementary Materials.
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
2
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 3

Table 1
Performances of best representatives of clustered pharmacophore hypotheses generated for PPARg activators.
Training
seta
Runb
HypothesesPharmacophoric features in
generated hypotheses
Total cost Cost of null
hypothesis
Residual
costc
Rd
F-Statistice
Cat-scramble
(%)
I17HBA, 2?Hbic, NegIon, RingArom
HBA, 2?Hbic, NegIon, RingArom
2?HBA, NegIons, RingArom
2?HBA, 2?Hbic, NegIons
HBA, 33Hbic, NegIons
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
HBA, NegIons, 2?RingArom
HBA, 3?Hbic, RingArom
23HBA, Hbic, RingArom
HBA, 3?Hbic, RingArom
HBA, 3?Hbic, RingArom
HBA, 3?Hbic, RingArom
HBA, 3?Hbic, RingArom
HBA, 3?Hbic, RingArom
HBA, 3?Hbic, RingArom
2?HBA, 2?Hbic, NegIons
2?HBA, 2?Hbic, NegIons
2?HBA, 2?Hbic, NegIons
3?HBA, NegIons
2?HBA, 2?Hbic, NegIons
2?HBA, 2?Hbic, NegIons
2?HBA, 2?Hbic, NegIons
3?HBA, NegIons
2?HBA, 3?Hbic
HBA, 3?Hbic, NegIons
2?HBA, 2?Hbic, NegIons
3?HBA, NegIons
2?HBA, 2?Hbic, NegIons
2?HBA, 3?Hbic
3?HBA, Hbic
3?HBA, Hbic
2?HBA, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
23HBA, 23Hbic, NegIons
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
2?HBA, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
2?HBA, 3Hbic
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
HBA, 3?Hbic, NegIons
HBA, 3?Hbic, NegIons
HBA, 2?Hbic, NegIons, RingArom
HBA, 3?Hbic, NegIons
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
2?HBA, 2?Hbic, NegIons
HBA, 3?Hbic, Negions
HBA, 2?Hbic, NegIons, RingArom
HBA, 4xHbic
HBA, 2?Hbic, NegIons, RingArom
HBA, 3?Hbic, RingArom
HBA, 3?Hbic, RingArom
HBA, Hbic, 2?RingArom
HBA, 3?Hbic, RingArom
2?HBA, 3?Hbic
2?Hbic, NegIons, 2?RingArom
HBA, 3?Hbic, RingArom
HBA, 4xHydrophobic
3?Hbic, 2?RingArom
2?Hbic, NegIons, 2?RingArom
2?Hbic, NegIons, 2?RingArom
97.7
98.1
89.8
95.4
89.4
91
96.5
97.1
91.2
84.1
100.5
82.8
94.3
97.2
92.6
94.9
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
111.4
13.6
13.2
21.6
15.9
22
20.4
14.9
14.3
20.1
27.2
10.9
28.5
17.1
14.2
18.8
16.5
0.83
0.83
0.94
0.86
0.92
0.89
0.84
0.83
0.93
0.89
0.8
0.77
0.90
0.86
0.901
0.88
23.9
65.4
52.3
61.6
83.6
72.4
49.7
49.2
60.8
97.3
43.2
35.1
58.8
32.5
63.6
63.8
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
10
21
4
1f
4
8
3
4
10
5
2
8f
26
10
73
10
83
10
II12
8
1
3
1
9
2
5
7
135
139.8
130.1
132.6
128.9
138.5
131.7
136.4
140.8
141.2
138.8
139.6
140.8
141.5
139
140.3
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
166.1
31.2
26.3
36
33.5
37.3
27.6
34.4
29.7
25.4
25
27.4
26.5
25.3
24.6
27.1
25.8
0.87
0.83
0.92
0.90
0.91
0.83
0.90
0.85
0.84
0.84
0.85
0.84
0.82
0.82
0.86
0.84
28.4
54.9
61.7
60.5
46.4
46.0
42.9
44.0
59.0
76.1
32.8
30.2
43.9
67.1
91.4
54.5
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
2
3
4
5
10
63
7
5
8
4
6
7
8
III18 94.6
94.8
90.7
93.0
90.1
91.7
93.7
94.3
95.2
95.6
95.9
96.0
91.7
95.6
95.5
95.6
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
124.4
29.8
29.6
33.7
31.4
34.4
32.8
30.8
30.4
29.3
28.9
28.6
28.4
32.8
28.8
29.0
28.8
0.93
0.93
0.97
0.94
0.95
0.95
0.93
0.92
0.92
0.92
0.93
0.92
0.97
0.93
0.94
0.94
68.0
72.4
75.1
62.9
32.9
57.5
73.5
50.1
33.4
48.5
19.3
70.6
56.6
27.9
66.6
73.1
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
10
22f
7
1
9
3
6
8
3
4
5
10
62
5
67
10
83
4
IV12
6
1
87.5
89.7
87.3
94.4
90.2
91.2
93.3
95.9
92.5
92.6
92.1
92.6
92.1
94.9
92.1
95.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
113.2
25.7
23.5
25.9
18.8
23.0
22.0
19.9
17.3
20.8
20.6
21.1
20.6
21.2
18.3
21.1
18.0
0.93
0.91
0.94
0.86
0.92
0.91
0.89
0.86
0.92
0.92
0.91
0.90
0.92
0.91
0.75
0.90
60.7
55.9
46.2
19.8
54.2
63.1
40.0
41.9
83.1
42.9
28.8
53.3
38.5
36.6
82.0
35.7
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
2
10
32
8
3
9
3
4
5
7
1
7
2
6
4
5
6
7
8
V15
7
3
9
2
77.8
78.2
78
79.4
76.7
78.1
104.6
104.6
104.6
104.6
104.6
104.6
26.8
26.4
26.6
25.3
27.9
26.5
0.97
0.97
0.96
0.94
0.97
0.96
66.9
57.4
37.6
33.3
55.0
64.6
90
90
90
90
90
90
2
3
10
(continued on next page)
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
3
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 4

However, although considered the most reliable structural
information that can be used for drug design, crystallographic
structures are limited by inadequate resolution [48] and crystalli-
zation-related artifacts of the ligandeprotein complex [49e51].
Moreover, crystallographic structures generally ignore structural
heterogeneity related to protein anisotropic motion and discrete
conformational substrates particularly in cases of pronounced
induced-fit protein flexibilities [52].
ThecontinuedinterestindesigningnewPPARgagonits,combined
with the drawbacks of structure-based design and the significant
induced-fit flexibility observed for PPARg [53], prompted us to
explorethepossibilityofdevelopingligand-basedthree-dimensional
(3D) pharmacophore(s) integrated within self-consistent QSAR
model(s). This approach avoids the pitfalls of structure-based tech-
niques;furthermore,thepharmacophoremodel(s) canbeusedas3D
templates to synthesize new PPARg agonists scaffolds.
We previously reported the use of this approach towards the
discovery of new inhibitory leads against glycogen synthase kinase
3b [54], hormone sensitive lipase [55], bacterial MurF [56], protein
tyrosine phosphatase 1B [57], influenza neuraminidase [58],
estrogen receptor beta ligands [59], cholesteryl ester transfer
protein inhibitors [60] and b-secretase inhibitors [61], CDK1
inhibitors [62], and Heat Shock Protein 90a Inhibitors [63].
We employed the CATALYST-HYPOGEN module embedded in
Discovery Studio (version 2.5) [64] to construct numerous
reasonable binding hypotheses for PPARg agonists. Subsequently,
genetic function algorithm (GFA) and multiple linear regression
(MLR) analyses were employed to search for an optimal QSAR that
combine high-quality binding pharmacophores with other molec-
ular descriptors and capable of explaining bioactivity variation
across a collection of diverse PPARg agonists. Optimal pharmaco-
phores were validated by evaluating their abilities to successfully
classify a list of compounds as actives or inactives byassessing their
receiver operating characteristic (ROC) curves.
2. Results and discussion
CATALYST-HYPOGEN models drugereceptor interactions using
information derived only from the drug structure [54e63]. It
requires a collection of training molecules with bioactivities
ranging from 3 to 4 orders of magnitude to attempt explain
bioactivity variation with respect to geometric localization of
chemical features in training molecules. To do this, it identifies
a 3D array of a maximum of five chemical features common to
active training molecules, which provides a relative alignment for
each input molecule consistent with their binding to a proposed
common receptor site. The chemical features considered can be
hydrogen bond donors and acceptors (HBD and HBA), aliphatic and
aromatic hydrophobes (Hbic), positive and negative ionizable
(PosIon and NegIon) groups and aromatic planes (RingArom). The
Table 1 (continued)
Training
seta
Runb
HypothesesPharmacophoric features in
generated hypotheses
Total costCost of null
hypothesis
Residual
costc
Rd
F-Statistice
Cat-scramble
(%)
44
8
2
HBA, 3?Hbic, RingArom
4xHbic, NegIons
3Hbic, NegIons, RingArom
2?Hbic, NegIons, 2?RingArom
3?Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
3?Hbic, NegIons, RingArom
3Hbic, NegIons, RingArom
HBA, 2?Hbic, NegIons, RingArom
2?Hbic, NegIons, 2?RingArom
2?HBA, Hbic, RingArom
HBA, 2?Hbic, RingArom
2?HBA, Hbic, RingArom
HBA, Hbic, 2?RingArom
HBA, HBD, 2?Hbic
HBA, HBD, Hbic, RingArom
2HBA, HBAD, Hbic
2?HBA, HBD, Hbic
HBA, 3?Hbic, RingArom
2?HBA, 3?Hbic
HBA, Hbic, 2?RingArom
2?HBA, 2?Hbic
HBA, 2?Hbic, RingArom
HBA, 3Hbic
2?HBA, 3?Hbic
2?HBA, 3?Hbic
2?HBA, 3?Hbic
2?HBA, 3?Hbic
2?HBA, HBD, Hbic
2?HBA, HBD, Hbic
2?HBA, HBD, Hbic
HBA, HBD, 3?Hbic
HBA, HBD, 3xHbic
HBA, HBD, 3?Hbic
77.9
78
75.9
76.7
75.9
76.6
77.1
77.3
76
76.3
104.6
104.6
104.6
104.6
104.6
104.6
104.6
104.6
104.6
104.6
26.7
26.6
28.8
27.9
28.7
28
27.5
27.3
28.7
28.3
0.97
0.97
0.98
0.96
0.97
0.96
0.96
0.95
0.96
0.96
88.4
42.2
13.3
61.2
46.6
73.3
18.1
25.2
44.2
51.1
90
90
90
90
90
90
90
90
90
90
5
10
62
4
5
7
6
9
7
8
VI11
8
6
95.9
99.6
100
101.1
99.5
98.5
97.1
99.8
101
101
101
101
101
101
101
101
5.1
1.4
1
0.1
1.5
2.5
3.9
1.2
0.91
0.83
0.83
0.81
0.87
0.76
0.84
0.82
10.8
72.7
80.5
6.2
0.2
0.2
2.1
2.5
90
90
90
90
90
90
90
90
4
10
61
10
82
10
VII12
4
3
6
2
4
1
6
3
84.5
79
89.5
91.7
90.1
91
87.3
94
88.6
95.5
92.9
95.3
90.1
93.4
88.4
89.8
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
121.3
36.7
42.3
31.7
29.6
31.1
30.2
33.9
27.3
32.6
25.7
28.4
25.9
31.1
27.9
32.8
31.5
0.96
0.95
0.93
0.91
0.94
0.93
0.94
0.88
0.88
0.82
0.89
0.87
0.92
0.88
0.91
0.90
0.0
0.6
2.5
1.2
1.0
3.1
8.1
6.9
0.4
1.8
0.0
2.7
1.4
4.4
3.7
3.9
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
2
3
4
5
10
62
6
1
5
4
9
7
8
aCorrespond to training sets in Table B under Supplementary Materials.
bCorrespond to runs in Table C under Supplementary Materials.
cDifference between total cost and the cost of the corresponding null hypotheses.
dCorrelation coefficients between pharmacophore-based bioactivity estimates and bioactivities of corresponding training compounds (subsets in Table B under
Supplementary Material).
eFisher statistic calculated based on the linear regression between the fit values of all collected inhibitors (1e88, Table A under Supplementary Material) against phar-
macophore hypothesis (employing the “best fit” option and Eq. (D) under Supplementary Material) and their respective PPARg activators bioactivities (log (1/IC50) values).
fBolded pharmacophores appeared in the best QSAR equations.
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
4
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 5

conformational flexibility of training ligands is modeled by
creating multiple conformers, judiciously prepared to emphasize
representative coverage over a specified energy range. CATALYST
pharmacophores have been used as 3D queries for mining struc-
tural databases for new active leads [54e63,65].
In the present project, we generated diverse hypotheses for
a series of PPARg activators. A total of 88 compounds were used in
this study (Fig. 1 and Table A under Supplementary Material)
[16,18e20,66,67]. Seven training subsets were selected from the
collection (Table B under Supplementary Material). Each subset
consisted of inhibitors of wide structural diversity.
2.1. Exploration of PPARg pharmacophoric space
The literature was surveyed to collect as many reported struc-
turally diverse PPARg activators as possible (1e88, see Fig. 1 and
Table A under Supplementary Material) [16,18e20,66,67]. The 2D
structures of the agonists were imported into Discovery Studio 2.5
and converted automatically into 3D single conformer represen-
tations. The structures were used as starting points for conforma-
tional analyses and in the determination of various molecular
descriptors for QSAR modeling.
The conformational space of each agonist was extensively
sampled utilizing the poling algorithm employed within Discovery
Studio2.5 [64].Conformational coveragewas performed
employing the “Best” module to ensure extensive sampling of
conformational space.
Subsequently,CATALYST-HYPOGEN
Studio was employed to identify as many pharmacophoric binding
modes assumed by PPARg agonists as possible. HYPOGEN imple-
ments an optimization algorithm that evaluates large number of
potential binding models for a particular target through fine
perturbations to hypotheses that survived the constructive and
subtractive phases of the modeling algorithm (see Section 4.1.4
Pharmacophoric Hypotheses Generation in Experimental and
Supplementary Materials) [68]. The extent of the evaluated
pharmacophoric space is reflected by the configuration (Config.)
cost calculated for each modeling run. It is generally recom-
mended that the Config. cost of any HYPOGEN run not to exceed 17
(corresponding to 2 [48] hypotheses to be assessed by CATALYST)
to guarantee thorough analysis of all models [69]. The size of the
investigated pharmacophoric space is a function of training
compounds, selected input chemical features and other CATALYST
control parameters [69].
Restricting the size of explored pharmacophoric space should
improve the efficiency of optimization via allowing efficient
assessment of limited number of pharmacophoric models. On the
other hand, extreme restrictions imposed on the evaluated phar-
macophoric space might reduce the possibility of discovering
optimal binding hypotheses, as they might occur outside the
“boundaries” of the evaluated space.
Therefore, it was decided to explore the pharmacophoric space
of PPARg agonists under reasonably imposed “boundaries” through
56 HYPOGEN automatic runs and employing seven carefully
selected trainingsubsets: subsets
Supplementary Material. The
subsets were selected in such away to guarantee maximal 3D
diversity and continuous bioactivity spread over more than 3.5
logarithmic cycles. We gave special emphasis to the 3D diversity of
the most active compounds in each training subset (Table B under
Supplementary Material) because of their significant influence on
the extent of the evaluated pharmacophoric space during the
constructive phase of HYPOGEN algorithm (see Section 4.1.4
Pharmacophoric Hypotheses Generation in Experimental and
Supplementary Materials) [68].
Guided by our rationally restricted pharmacophoric exploration
concept, we restricted the software to explore pharmacophoric
models incorporating from zero to one NegIon feature, from zero to
three HBA, Hbic, and RingArom features instead of the default
range of 0e5, as shown in Table C under Supplementary Material.
Furthermore, we instructed HYPOGEN to explore only 4- and 5-
featured pharmacophores, i.e., ignore models of lesser number of
features in order to further narrow the investigated pharmaco-
phoric space and to better represent the diverse interactions
between known ligands and PPARg binding pocket (as shown in
Table C under Supplementary Material).
Ineachrun,theresultingbindinghypotheseswereautomatically
ranked according to their corresponding “total cost” value, which is
defined as the sum of error cost, weight cost and configuration cost
(see Section 4.1.5) [68e72]. Error cost provides the highest contri-
bution to total cost and it is directly related to the capacity of the
particularpharmacophoreas3D-QSARmodel,i.e., incorrelating the
molecular structures to the corresponding biological responses
[68e72]. HYPOGEN also calculates the cost of the null hypothesis,
which presumes that there is no relationship in the data and that
experimental activities are normally distributed about their mean.
Accordingly,thegreaterthedifferencefromthenullhypothesiscost
(residual cost, Table 1) the more likely that the hypothesis does not
reflect a chance correlation. An additional validation technique
based on Fisher’s randomization test [73] was recently introduced
modulein Discovery
1e7 inTableB
in
under
thesetraining compounds
A
-4
-3
-2
-1
0
1
2
3
4
-4 -3 -2-101234
Experimental log (1/IC50)
C
I / 1 (
g
o l
d
e t c i d
e r
P
0
5)
B
-4
-3
-2
-1
0
1
2
3
4
-4-3 -2-101234
Experimental log(1/IC50)
C
I / 1 ( g
o l
d
e t c i d
e r
P
0
5)
Fig. 2. Experimental versus (A) fitted (71 compounds, rLOO2¼0.73), and (B) predicted
(17 compounds, rPRESS2¼ 0.67) bioactivities calculated from the best QSAR model Eq.
(1). The solid lines are the regression lines for the fitted and predicted bioactivities of
training and test compounds, respectively, whereas the dotted lines indicate 1.0 log
point error margins.
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
5
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 6

intoCATALYST:Cat.Scramble[64].Inthistestthebiologicaldataand
the corresponding structures are scrambled several times and the
softwareischallengedtogeneratepharmacophoricmodelsfromthe
randomized data. The confidence in the parent hypotheses (i.e.,
generated from unscrambled data) is lowered proportional to the
number of times the software succeeds in generating binding
hypotheses from scrambled data of apparently better cost criteria
than the parent hypotheses (see Section 4.1.5) [68e72].
Eventually, 519 pharmacophore models emerged from 56
automatic HYPOGEN runs, out of which only 500 models illustrated
Cat. scramble confidence levels?90%. These successful models
were clustered and the best representatives (104 models) were
used in subsequent QSAR modeling (see Section 4.1.6). Table 1
shows the statistical criteria of representative cluster centers.
Clearly, from the table, representative models shared comparable
features and acceptable statistical success criteria.
The fact that many pharmacophore models were optimal and
statistically comparable suggests the ability of PPARg ligands to
assume multiple pharmacophoric binding modes within the
binding pocket. Therefore, it is quite challenging to select any
Fig. 3. Pharmacophoric features of (A) S1R3H1, (C) S1R5H8, (E) S3R2H2, (B) S1R3H1fitted against 10 (IC50¼1 nM) (D) S1R5H8fitted against 49 (IC50¼19 nM), (F) S3R2H8fitted against
53 (IC50¼30 nM). HBA shown as green vectored spheres, Hbic as light blue spheres, RingArom as vectored orange spheres and NegIon as dark blue sphere. (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version of this article.)
Table 2
Thecross-correlationr2betweenthesuccessfulpharmacophoreshypothesesinEq.(1).
S1R3H1
1
0.74
0.25
S3R2H2
S1R5H8
S1R3H1
S3R2H2
S1R5H8
1
0.201
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
6
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 7

particular pharmacophore hypothesis as a sole representative of
the binding process.
2.2. QSAR modeling
Despite that pharmacophoric hypotheses provide excellent
insights into ligand-macromolecule recognition and can be used to
mine for new biologically interesting scaffolds, their predictive
value as 3D-QSAR models is limited by steric shielding and bioac-
tivity-enhancing or reducing auxiliary groups [54e58,60,61,74].
This point combined with the fact that pharmacophore modeling of
PPARg activators furnished several binding hypotheses of compa-
rable success criteria prompted us to employ classical QSAR anal-
ysis to search for the best combination of pharmacophore(s) and
other 2D descriptors capable of explaining bioactivity variation
across the whole list of collected inhibitors (1e88, Table A in
Supplementary Materials and Fig. 1). We employed genetic func-
tion approximation and multiple linear regression QSAR (GFA-
MLR-QSAR) analysis to search for an optimal QSAR equation(s).
The fit values obtained by mapping representative hypotheses
(104 models) against collected PPARg activators (1e88) were
enrolled, togetherwithnearly
descriptors, as independent variables (genes) in GFA-MLR-QSAR
analysis (see Section 4.1.7) [54e58,60,61,74,75]. However, since it is
essential to access the predictive power of the resulting QSAR
models on an external set of inhibitors, we randomly selected 17
molecules (marked with double asterisks in Table A, see Section
4.1.7) and employed them as external test molecules for validating
the QSAR models (rPRESS2). Moreover, all QSAR models were cross-
validated automatically using the leave-one-out cross-validation in
Discovery Studio 2.5 [64].
Eq. (1) shows the details of the optimal QSAR model. Fig. 2
shows the corresponding scatter plots of experimental versus
estimated bioactivities for the training and testing inhibitors.
100 otherphysicochemical
where, r712is the correlation coefficient against 71 training
compounds, rLOO2is the leave-one-out correlation coefficient, radj2
is r2adjustedfor the numberof terms in the model and rPRESS2is the
predictive r2determined for the 17 test compounds [64,75]. S1R3H1
and S1R5H8 represent the fit values of the training compounds
against the first and eighth pharmacophoric Hypotheses generated
in third and fifth HYPOGEN Runs, respectively, using the first
training Subset, while S3R2H2represents the fit values against the
second pharmacophoric Hypothesis generated during from the
second HYPOGEN Run performed on the third training Subset.
Table C (under Supplementary Materials) and Table 1 show the
pharmacophore modeling runs and the statistical criteria of output
models. Bolded runs in Table 1 correspond to the three QSAR-
selected pharmacophores (i.e., S1R3H1, S1R5H8and S3R2H2). The fit
values were calculated based on Eq. (D) under Supplementary
Materials.
Fig. 3 shows S1R3H1, S1R5H8and S3R2H2and how they map three
potent trainingcompounds,
(IC50¼19 nM), 53 (IC50¼30 nM), while Table 3 showsthe X,Y, and Z
coordinates of the three pharmacophores.
HBD is the number of hydrogen bond donors in a particular
molecule, RotableBonds is the number of rotable bonds, FPSA is the
molecular fractional polar surface area (calculated as the ratio of
the polar surface area divided by the total surface area), and LUMO
is the energy of the lowest unoccupied molecular orbital calculated
employing the density functional theory method implemented in
DS 2.5 [64].
Emergence of three orthogonal pharmacophoric models, i.e.,
S1R3H1, S1R5H8and S3R2H2(of cross-correlation r2?0.74, Table 2)
in Eq. (1) suggests that they represent three complementary
binding modes accessible to ligands within the binding pocket of
PPARg, which means thatone of the pharmacophores can optimally
explain the bioactivities of some training inhibitors, while the
others explain the remaining inhibitors. Similar conclusions were
namely,
10
(IC50¼1 nM),
49
Table 3
Pharmacophoric features and corresponding weights, tolerances and 3D coordinates of S1R3H1, S1R5H8and S3R2H2.
Model DefinitionsChemical features
HBAHbic HbicHbicNegIons
S1R3H1a
Weights
Tolerances
Coordinates
2.29089
1.60
2.79
?4.10
0.46
2.29089
1.60
1.18
?3.57
7.38
2.29089
1.60
?0.38
0.77
?6.1
Hbic
2.29089
1.60
?1.61
?1.65
1.43
2.29089
1.60
?2.71
6.3
1.11
2.20
3.17
?6.41
?1.42
X
Y
Z
HBAHBARingArom
S1R5H8b
Weights
Tolerances
Coordinates
2.57353
1.60
3.09
?4.19
0.33
2.57353
1.60
1.19
?0.53
?3.23
NegIons
2.57353
1.60
0.33
3.62
0.13
2.57353
1.60
4.84
?1.90
2.62
2.20
1.69
?5.19
?2.13
Hbic
2.20
?0.55
1.90
?3.57
HBA
1.60
6.41
?0.44
0.52
X
Y
Z
Hbic HBA
S3R2H2c
Weights
Tolerances
Coordinates
2.42026
1.60
?0.27
?5.46
?1.94
2.42026
1.60
12.92
?0.84
?4.23
2.42026
1.60
9.44
?2.87
1.57
2.42026
1.6
7.65
?4.41
3.25
2.42026
1.6
5.83
?0.06
?4.92
2.20
6.30
?4.25
5.92
2.20
4.16
1.15
?7.10
X
Y
Z
aS1R3H1: the code refer to training Subset I (see Table B in Supplementary Materials), third Run and first pharmacophore Hypothesis, as in Table 1 and Table C under
Supplementary Materials.
bThe 8th Hypothesis from 5th Run on 1st training Subset.
cThe 2nd Hypothesis from 2nd Run on 3rd Subset.
logð1=IC50Þ ¼ ?2:4 þ 0:30ðS1R5H8Þ þ 5:63 ? 10?2ðS3R2H2Þ þ 9:76 ? 10?2ðS1R3H1Þ
þ5:1FPSA ? 9:3LUMO ? 0:05ðRotable BondsÞ ? 0:28HBD
r2
71¼ 0:80;F ? statistic ¼ 270:3;r2
adj¼ 0:77;r2
LOO¼ 0:73;r2
PRESSð17Þ¼ 0:67
ð1Þ
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
7
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 8

reached about the binding pockets of other targets based on QSAR
analysis [54e58,60,61,74].
Emergence of FPSA in Eq. (1) in association with positive slope
suggests a direct relationship between ligand/PPARg affinity and
ligands’ hydrophilicity. We believe this trend is explainable by the
fact that hydrophilic ligands favor docking into the binding site due
to the presence of good number of hydrophilic amino acid residues
in binding site (e.g., Lys367, Ser289, Arg288, Glu286, His449,
His323, Tyr473, Gln283, and Asp362). However, this seems to
contradict with the appearance of HBD in Eq. (1) combined with
negative slope, which suggest that ligands of more hydrogen bond
donors disfavor binding. Nevertheless, careful evaluation of
training compounds reveals that excess HBDs in poorly potent
ligands are always negative ionizable, thus they render their cor-
responding ligands extremely hydrophilic and favor hydration
instead of binding within bindng pocket.
On the other hand, emergence of LUMO in Eq. (1) combined
with negative slope suggests that ligand/PPARg affinity favors
Fig. 4. X-ray structures of three ligands co-crystallized within PPARg binding pocket: (A) 2Q59 (resolution¼2.20 Å), (D) 2G0 G (resolution¼ 2.54 Å) and (G) 2P4Y (reso-
lution ¼2.25 Å). (B), (E) and (H) mapping pharmacophores S1R3H1, S1R5H8, and S3R2H2against the co-crystallized ligands of 2Q59, 2G0G, and 2P4Y (rigid mapping), respectively; (C),
(F) and (I) the chemical structures of the co-crystallized ligands of 2Q59, 2G0G, and 2P4Y, respectively.
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
8
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 9

electrophilic ligands (i.e., of more negative LUMO values) probably
due top-stacking with certain electron-rich aromatic centers in the
binding pocket (e.g., Phe363, Phe287, Tyr327, and Tyr473).
Finally, appearance of RotableBonds in Eq. (1) combined with
negative regression coefficient suggests that flexible molecules
disfavor binding with PPARg, which is not unexpected since the
entropic costof protein binding with flexibleligands is muchhigher
that that required for rigid molecules. Therefore, a significant frac-
tion of energy gains resulting from binding enthalpy will be wasted
on entropic costs required for fixing flexible molecules within the
binding pocket.
2.3. Comparing pharmacophore models with crystallographic
complexes
To further emphasize the validity of our QSAR-selected phar-
macophores, we compared the crystallographic structures of three
PPARg/ligand complexes (PDB codes: 2Q59, 2G0G, and 2P4Y) with
S1R3H1, S1R5H8, and S3R2H2. Fig. 4 shows the chemical structures of
the ligands and compares their PPARg complexes with the ways
they map S1R3H1, S1R5H8, and S3R2H2employing rigid mapping, i.e.,
fitting the ligands’ bound states against corresponding pharmaco-
phores without conformational adjustments.
Fitting the carboxylate of the 2Q59 ligand against NegIon in
S1R3H1(Fig.4b)correspondstoelectrostaticinteractionsconnecting
this fragment with the imidazole side chains of His449 and His323,
as in Fig. 4a. Similarly, mapping the ligand’s amidic carbonyl with
Table 4
ROC curve analysis criteria for QSAR-selected pharmacophores and their sterically
refined versions.
Pharmacophore modelROCaeAUCb
ACCc
SPCd
TPRe
FNRf
S1R3H1
S1R5H8
S3R2H2
0.99
0.82
0.97
0.96
0.96
0.96
0.97
0.95
0.99
0.56
1.00
0.28
0.026
0.047
0.013
aROC: receiver operating characteristic curve.
bAUC: area under the curve.
cACC: overall accuracy.
dSPC: overall specificity.
eTPR: overall true positive rate.
fFNR: overall false negative rate.
A
0 0.10.2 0.30.4 0.50.6 0.70.8 0.91
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FALSE POSITIVE RATE
TRUE POSITIVE RATE
RECEIVER OPERATING CHARACTERISTIC (ROC)
B
0 0.10.2 0.30.40.5 0.60.7 0.80.91
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FALSE POSITIVE RATE
TRUE POSITIVE RATE
RECEIVER OPERATING CHARACTERISTIC (ROC)
C
00.1 0.20.30.4 0.50.6 0.70.8 0.91
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FALSE POSITIVE RATE
TRUE POSITIVE RATE
RECEIVER OPERATING CHARACTERISTIC (ROC)
Fig. 5. ROC curves of: (A) S1R3H1, (B) S3R2H2, (C) S1R5H8.
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
9
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 10

a HBA feature in S1R3H1(Fig. 4b) corresponds to hydrogen bonding
interaction between this carbonyl and the explicit water molecule
H2O35. Apparently, this water molecule is fixed via hydrogen
bonding with guanidine of Arg288 (Fig. 4a). Similarly, mapping the
trifluoromethoxy of 2Q59 ligand against Hbic feature in S1R3H1
(Fig. 4b) correlateswith fittingthis group intoa hydrophobic pocket
composed of the hydrophobic side chains of Met364, Ile281, Le353
and Met348, as in Fig. 4a. Finally, mapping the indole methyl and
O
O
O
OH
OH
O
O
OH
OH
O
O
O
NH
O
O
O
NH
O
O
O
N
N
N
N
S
OH
O
OH
O
N
N
OH
O
NH
OH
S
O
OH
O
89
901929
S
N
O
O
O
OH
O
S
O
O
NH
S
S
NH
O
OOH
OH
O
O
N+
OH
O
OO
O
O
93*
94
95*
N
N
N
S
O
OH
O
O
N
N
O
O
O
NH
O
N
O
H2
H
OH
O
O
O
O
N
N
S
O
O
O
OH
O
O
*96 9789
O
S
S
O
OH
O
N
O
S
S
O
OH
O
N
O
CH3
99001
O
S
S
O
OH
O
N
Cl
Cl
101
Fig. 6. Chemical structure of captured hits. Active hits are marked with asterisks. Corresponding bioactivities are shown in Table 5.
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
10
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 11

aromatic methoxy of the co-crystallized ligand against two Hbic
features in S1R3H1(Fig. 4b) agrees with positioning these groups at
close proximity with the hydrophobic side chains of Ile326 and
Ile296, respectively, in the bound crystallographic structure of the
ligand (Fig. 4a).
A similar trend can be seen by comparing the crystallographic
structure of bound ligand in 2G0 G with the way it maps S1R5H8
(Fig. 4d and e). Mapping the ligand’s pyrazole nitrogen and sulfo-
namidic oxygen against two HBAs in S1R5H8 corresponds to
hydrogen bonding interactions connecting these atoms with SH
and OH side chains of Cys285 and Ser289, respectively, as in Fig. 4d
and e. On the other hand, mapping the fluorobenzene and thio-
phene rings of the bound ligand against RingArom and Hbic
features in S1R5H8, respectively (Fig. 4e), correlates with stacking
these rings against the hydrophobic side chains of Leu330 and
Ile281 in 2G0G (Fig. 4d).
Finally, S3R2H2 seems to match the co-crystallized complex
2P4Y (Fig. 4g and h): Mapping the ligand’s carboxylate against HBA
and NegIon features in S3R2H2corresponds to hydrogen bonding
and electrostatic attraction connecting this carboxylate with the
OH and guanidine side chains of Ser342 and Arg288 in 2P4Y.
Similarly, mapping the methoxy benzene ring against Hbic feature
in S3R2H2 correlates with projecting this group close to the
hydrophobicside chainsofMet329
The hydrogen bonding interaction between the OH of Ser289 and
the ether oxygen of trifluoromethoxy of 2P4Y ligand agrees with
mapping this group against HBA feature in S3R2H2. Finally, the close
proximity between the chloro- substituent of the ligand’s central
aromatic linker correlates with placing this group in a hydrophobic
pocket comprised of the side chains of Met348, Leu353 and
Met364.
Clearly from the above discussion, the three pharmacophores
S1R3H1, S1R5H8, and S3R2H2represent three valid binding modes
assumed by ligands within PPARg. Incidentally, the three phar-
macophores point to limited number of critical interactions
required for high ligand-PPARg affinity in each of the binding
modes. In contrast, crystallographic complexes reveal many
bonding interactions without highlighting critical ones. Fig. 4a,
d and g shows only interactions corresponding to pharmacophoric
features while other binding interactions were hidden for clarity.
andAla292in 2P4Y.
2.4. Receiver operating characteristic (ROC) curve analysis
To further validate the resulting models (both QSAR and phar-
macophores), we subjected our QSAR-selected pharmacophores to
receiver operating curve (ROC) analysis. In ROC analysis, the ability
of a particular pharmacophore model to correctly classify a list of
compounds as actives or inactives is indicated bythe area under the
curve (AUC) of the corresponding ROC as well as other parameters,
namely, overall accuracy, overall specificity, overall true positive
rate and overall false negative rate (see Section 4.1.8 for more
details) [76e79].
Table 4 and Fig. 5 show the ROC results of our QSAR-selected
pharmacophores. S1R3H1and S3R2H2illustrated excellent overall
performances with ROC-AUC values exceeding 95%. On the other
hand, S1R5H8exhibited moderate performance with AUC value of
82%. This is not unexpected, as both S1R3H1and S3R2H2are 5-
featured pharmacophores, while S1R5H8exhibit 4 features only.
Higher-featured pharmacophores are expected to be more selective
as 3D search queries since additional features impose more provi-
sions on captured hits.
2.5. In silico screening and subsequent in vitro evaluation
S1R3H1, S1R5H8and S3R2H2were employed as 3D search queries
against the National Cancer Institute list of compounds (NCI,
238,819 structures) using the “Best Flexible Database Search”
option implemented within CATALYST. Compounds that have their
chemical groups spatially overlap (map) with corresponding
features of each particular pharmacophoric model were captured
as hits. Captured hits were filtered based on Lipinski’s and Veber’s
rules [80,81]. Surviving hits were fitted against S1R3H1, S1R5H8,
S3R2H2and their fit values, together with other relevant molecular
descriptors, were substituted in QSAR Eq. (1) to predict their
affinity IC50
values. The highest-ranking available hits (13
compounds, Fig. 6) were evaluated in vitro for potential PPARg
ligand activity using a PPRE-luciferase reporter system transfected
in HepG2 cells. Hits were initially screened at 40 mM concentra-
tions, subsequently; compounds that significantly activated PPARg
were further assessed to determine their EC50values. Table 5 lists
hits that illustrated significant PPARg ligand activities; estimated
affinities to PPARg and in vitro experimental bioactivation EC50
values.
To validate our bioassay settings we determined the EC50value
of the standard PPARg activator rosiglitazone under the same
conditions. Our conditions determined the EC50value of rosiglita-
zone to be 10 nM, which is within the reported literature range of 2
to 16 nM [82,83].
Clearly from Table 5, in vitro testing showed that 3 NCI high-
ranking hits activated PPARg at nanomolar to micromolar EC50
values. Fig. 6 shows the chemical structures of the tested hits
including active ones, while Fig. 7 shows how the most potent hit
93 fits the three successful pharmacophore models.
However, it must be mentioned that QSAR and pharmacophore
modeling were based on affinity values (IC50), and therefore, the
corresponding predictions were in IC50format. On the other hand,
since we implemented a functional bioassay that detects agonistic
bioactivities of captured hits, i.e., EC50values, explains the limited
number of active hits and the apparent differences between pre-
dicted and experimental bioactivities.
Table 5
Predicted and experimental bioactivities of high-ranking hit molecules.
Hitsa
Nameb
Best fit valuesc
Predicted affinity IC50(nM)d
Experimental EC50(nM)e,g
S1R3H1
3.59542
2.11611
0.475151
0
S1R5H8
8.42089
7.40627
4.59272
5.332
S3R2H2
7.34686
0
0
0
93
95
96
Rosiglitazone
NCI144248
NCI197178
NCI289920
N/A
2.4
3.6
15.3 (r2¼ 0.99)f
220? 103(r2¼ 0.98)f
224 (r2¼ 0.99)f
10 (r2¼ 0.95)f
224.8
3.5
aChemical structures shown in Fig. 6.
bNCI number.
cFit values calculated against respective hypotheses using Eq. (D) in Supplementary Materials.
dPredicted IC50nM according to QSAR Eq. (1).
eExperimental bioactivation (EC50values) determined in triplicates.
fValues in brackets represent the correlation coefficients of the corresponding dose-response line.
gThese values represent average results of triplicate measurements.
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
11
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 12

3. Conclusions
PPARg activators are currently considered as potential treat-
ments for diabetes and hyperglycemia. The pharmacophoric space
of PPARg activators was explored via eight diverse sets of activators
and using CATALYST-HYPOGEN module of Discovery Studio to
identify high-quality binding model(s). Subsequently, genetic algo-
rithm and multiple linear regression analysis were employed to
access optimal QSAR model capable of explaining PPARg activation
variation across 88 collected PPARg activators. Three orthogonal
pharmacophoric models emerged in the QSAR equation suggesting
the existence of at least three distinct binding modes accessible to
activating ligands within PPARg binding pocket. The QSARequation
and the associated pharmacophoric models were experimentally
validated by the identification of three nanomolar to micromolar
PPARgactivatorsretrievedviainsilicoscreening.Ourresultssuggest
that the combination of pharmacophoric exploration and QSAR
analysescanbeusefultoolforfindingnewdiversePPARgactivators.
4. Experimental
4.1. Molecular modeling
4.1.1. Software and hardware
The following software packages were utilized in the present
research.
? CS ChemDraw Ultra 6.0, Cambridge Soft Corp. (www.
cambridgesoft.com), USA.
? Discovery Studio 2.5, Accelrys Inc. (www.accelrys.com), USA.
Pharmacophore and QSAR modeling studies were performed
using Discovery Studio 2.5 suite from Accelrys Inc. (San Diego,
California, www.accelrys.com) installed on installed on a Core 2
Duo Pentium PC.
4.1.2. Dataset
The
Supplementary Materials) were collected from published litera-
ture [16,18e20,66,67]. The in-vitro bioactivities of the collected
agonists were expressed as the concentration of the test compound
that causes 50% displacement of a radio-labeled full agonist bound
to the receptor (IC50). The logarithm of measured IC50(nM) values
were used in pharmacophore modeling and QSAR analysis, thus
correlating the data linear to the free energy change. In one
instance the IC50was reported to be below 1 nM (compound 13,
Table A under Supplementary Materials), however, to allow proper
QSAR modeling we assumed that IC50equals 1 nM. The logarithmic
transformation of IC50values should minimize any potential errors
resulting from this assumption.
The two-dimensional (2D) chemical structures of the inhibitors
were sketched using ChemDraw Ultra and saved in MDL-molfile
format. Subsequently, they were imported into Discovery Studio,
converted into corresponding standard 3D structures and energy
minimized to the closest local minimum using the molecular
mechanics CHARMm force field implemented in CATALYST module
of Discover Studio. The resulting 3D structures were utilized as
starting conformers for CATALYST-based conformational analysis.
structuresof88PPARg
ligands(TableAunder
4.1.3. Conformational analysis
The conformational space of each agonist (1e88, Table 1) was
explored adopting the “best conformer generation” option within
CATALYST module of Discovery Studio, which is based on the
generalized CHARMm force field implemented in the program.
Conformational ensembles were generated with an energy
Fig. 7. Mapping active hit 93 (NCI144248, EC50¼15.3 nM) against (A) S1R3H1, (B)
S1R5H8, and (C) S3R2H8(see Table 5). HBA shown as green vectored spheres, Hbic as
light blue spheres, RingArom as vectored orange spheres and NegIon as dark blue
sphere. (For interpretation of the references to colour in this figure legend, the reader
is referred to the web version of this article.)
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
12
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 13

threshold of 20 kcal/mol from the local minimized structure and
a maximum limit of 250 conformers per molecule.
4.1.4. Pharmacophoric hypotheses generation
All 88 molecules with their associated conformational models
were regrouped into a spreadsheet. The biological data of the
inhibitors were reported with uncertainty values of 2 or 3, which
means that the actual bioactivity of a particular inhibitor is
assumed to be situated somewhere in intervals ranging from 1/2 to
2 or 1/3 to 3 times the reported bioactivity value of that inhibitor,
respectively(See TableCunder
[70,84,85]. The uncertainty value is of great impact on the qualities
of the resulting pharmacophores, as it controls the number of
training compounds within the “most potent category” (see Eq. (A)
under section SM-1 in Supplementary Materials).
Subsequently, seven structurally diverse training subsets were
carefullyselected from the collection for pharmacophore modeling.
Table B under Supplementary Material shows the selected subsets
and their member compounds. Typically, CATALYST requires
informative training sets that include at least 16 compounds of
evenly spread bioactivities over at least three and a half logarithmic
cycles. Lesser training lists could lead to chance correlation and
thus faulty models.
The selected training sets were utilized to conduct 56 modeling
runs to explore the pharmacophoric space of PPARg (Table C under
SupplementaryMaterial).The
altering interfeature spacing parameter (100 and 300 picometers),
the uncertainty value (2 or 3) and the maximum numberof allowed
features in the resulting pharmacophore hypotheses, i.e., they were
allowed to vary from 4 to 5 or from 5 to 5. Furthermore, some
features were fixed in some runs, i.e., NegIon was fixed by setting
the number of possible NegIon features in the resulting pharma-
cophore models to one, while other features were allowed to vary
during pharmacophore modeling, as shown in Table C under
Supplementary Material.
Pharmacophore modelingemploying
through three successive phases: the constructive phase, subtrac-
tive phase and optimization phase (see CATALYST Modeling Algo-
rithm in section SM-1 under Supplementary Materials) [64,68e72].
SupplementaryMaterials)
exploration processincluded
CATALYSTproceeds
4.1.5. Assessment of the generated hypotheses
When generating hypotheses, CATALYST attempts to minimize
a cost function consisting of three terms: Weight cost, Error cost
and Configuration cost (see CATALYST Cost Analysis in Assessment
of Generated Binding Hypotheses under Supplementary Materials).
An additional approach to assess the quality of CATALYST-HYPO-
GEN pharmacophores is to cross-validate them using the Cat-
Scramble program implemented in CATALYST. This validation
procedure is based on Fischer’s randomization test [73]. In this vali-
dation tes, we selected a 95% confidence level, which instruct CATA-
LYST to generate 19 random spreadsheets by the Cat-Scramble
command. Subsequently, CATALYST-HYPOGEN is challenged to use
these random spreadsheets to generate hypotheses using exactly the
same features and parameters used in generating the initial
unscrambled hypotheses. Success in generating pharmacophores of
comparable cost criteria to those produced by the original unscram-
bled data reduces the confidence in the training compounds and the
unscrambled original pharmacophore models [64,73,86]. Based on
Fischerrandomizationcriteria;allthepharmacophoresexceededthe
85% significance threshold, and therefore were considered fit for
subsequent processing (clustering and QSAR analyses).
4.1.6. Clustering of the generated pharmacophore hypotheses
The pharmacophore models were clustered into 104 groups
utilizing the hierarchical average linkage method available in
CATALYST. Subsequently, the highest-ranking representatives, as
judged based on their significance F-values, were selected to
represent their corresponding clusters in subsequent QSAR
modeling. Table 1 shows the pharmacophoric features and statis-
tical criteria of representative pharmacophores including their
pharmacophoric features, success criteria and differences from
corresponding null hypotheses. The table also shows the corre-
sponding Cat. Scramble confidence levels determined for each
representative pharmacophore.
4.1.7. QSAR modeling
A subset of 72 compounds from the total list of inhibitors (1e88,
TableAunder Supplementary
[16,18e20,66,67] was utilized as a training set for QSAR modeling;
the remaining 16 molecules (ca. 20% of the dataset) were employed
as an external test subset for validating the QSAR models. The test
molecules were selected as follows: the 88 agonists were ranked
according to their IC50values, and then every fifth compound was
selected for the test set starting from the high-potency end. This
selection considers the fact that the test molecules must represent
a range of biological activities similar to that of the training set. The
selected test inhibitors are marked with double asterisks in Table A
(Supplementary Material).
The logarithm of measured 1/IC50(mM) values was used in QSAR,
thus correlating the data linear to the free energy change. The
chemical structures of the inhibitors were imported into Discovery
Studio as standard 3D single conformer representations in SD
format. Subsequently, different descriptor groups were calculated
for each compound. The calculated descriptors included various
simple and valence connectivity indices, electro-topological state
indices and other molecular descriptors (e.g., logarithm of partition
coefficient, polarizability, dipole moment, molecular volume,
molecular weight, molecular surface area, etc.) [64]. The training
compounds were fitted (using the best fit option in CATALYST
module of Discovery Studio) [64] against the representative phar-
macophores, and their fit values were added as additional descrip-
tors.ThefitvalueforanycompoundisobtainedautomaticallyviaEq.
(D) (under Section SM-2 in Supplementary Materials) [64].
Genetic function approximation (GFA) was employed to search
for the best possible QSAR regression equation capable of corre-
lating the variations in biological activities of the training
compounds with variations in the generated descriptors, i.e.,
multiple linear regression modeling (MLR). GFA techniques rely on
the evolutionary operations of ‘‘crossover and mutation’’ to select
optimal combinations of descriptors (i.e., chromosomes) capable of
explaining bioactivity variation among training compounds from
a large pool of possible descriptor combinations, i.e., chromosomes
population. However, to avoid overwhelming GFA-MLR with large
number of poor descriptor populations, we removed lowest-vari-
ance descriptors (20%) prior to QSAR analysis.
Each chromosome is associated with a fitness value that reflects
how good it is compared to other solutions. The fitness function
employed herein is based on Friedman’s ‘lack-of-fit’ (LOF) [64].
Our preliminary diagnostic trials suggested the following
optimal GFA parameters: explore linear, quadratic and spline
equations at mating and mutation probabilities of 50%; population
size¼500; number of genetic iterations¼30,000 and lack-of-fit
(LOF) smoothness parameter¼1.0. However, to determine the
optimal number of explanatory terms (QSAR descriptors), it was
decided to scan and evaluate all possible QSAR models resulting
from 4 to 10 explanatory terms.
All QSAR models werevalidated employing leave-one-out cross-
validation (rLOO2), bootstrapping (radj2) and predictive r2(rPRESS2)
calculated from the test subsets. The predictive rPRESS2is defined as
in Eq. (2):
MaterialsandFig. 1)
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
13
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 14

r2
PRESS¼ SD ? PRESS=SD
where SD is the sum of the squared deviations between the bio-
logical activities of the test set and the mean activity of the training
set molecules, PRESS is the squared deviations between predicted
and actual activity values for every molecule in the test set.
(2)
4.1.8. Receiver operating characteristic (ROC) curve analysis
The optimal pharmacophore models (i.e., S1R3H1, S1R5H8and
S3R2H2) were validated by assessing their abilities to selectively
capture diverse PPARg active compounds from a large testing list of
actives and decoys.
The testing list was prepared as described by Verdonk and co-
workers [76,77]. Briefly, decoy compounds were selected based on
three basic one-dimensional (1D) properties that allow the
assessment of distance (D) between two molecules (e.g., i and j): (1)
the number of hydrogen-bond donors (NumHBD); (2) number of
hydrogen-bond acceptors (NumHBA) and (3) count of nonpolar
atoms (NP, defined as the summation of Cl, F, Br, I, S and C atoms in
a particular molecule). For each active compound in the test set, the
distance to the nearest other active compound is assessed by their
Euclidean distance (Eq. (3)):
The minimum distances are then averaged over all active
compounds (Dmin). Subsequently, for each active compound in the
test set, around 30 decoys were randomly chosen from the ZINC
database [78]. The decoys were selected in such a way that they did
not exceed
Dmin
distance from their corresponding
compound.
To diversify active members in the list, we excluded any active
compoundhavingzerodistance(D(i,j))fromotheractivecompound
(s) in the test set. Active testing compounds were defined as those
possessing PPARg affinities ranging from 0.6 nM to 6.0 mM. The test
set included 25 active compounds and 776 ZINC decoys.
The test list (801 compounds) was screened by each particular
pharmacophore employing the “Best flexible search” option
implemented in Discovery Studio, while the conformational spaces
of the compounds were generated employing the “Fast conforma-
tion generation option”
implemented
Compounds missing one or more features were discarded from the
hit list. In-silico hits were scored employing their fit values as
calculated by Eq. (D) in Supplementary Materials.
The ROC curve analysis describes the sensitivity (Se or true
positive rate, Eq. (4)) for any possible change in the number of
selected compounds (n) as a function of (1-Sp). Sp is defined as
specificity or true negative rate (Eq. (5)) [77,79].
active
inDiscoveryStudio.
Se ¼Number of Selected Actives
Total Number of Actives
¼
TP
TP þ FN
(4)
Sp ¼Number of Discarded Inactives
Total Number of Inactives
¼
TN
TN þ FP
(5)
where, TP is the number of active compounds captured by the
virtual screening method (true positives), FN is the number of
active compounds discarded by the virtual screening method, TN is
the number of discarded decoys (presumably inactives), while FP is
the number of captured decoys (presumably inactive) [77,79].
A ROC curve is plotted by setting the score (fit value) of the
highest scoring active molecule as the first threshold. Afterwards,
the number of decoys within this cutoff is counted and the
corresponding Se and Sp pair is calculated [77,79]. This calculation
is repeated for the active molecule with the second highest score
and so forth, until the scores of all actives are considered as
selection thresholds.
The ROC curve representing ideal distributions, where no
overlap between the scores of active molecules and decoys exists,
proceeds from the origin to the upper-left corner until all the
actives are retrieved and Se reaches the value of 1. In contrast to
that, the ROC curve for a set of actives and decoys with randomly
distributed scores tends towards the Se¼1?Sp line asymptotically
with increasing number of actives and decoys [77,79]. The success
of a particular virtual screening workflow can be judged from the
following criteria (shown in Table 4):
(1) Area under the ROC curve (AUC) [77,79,87]. In an optimal ROC
curve an AUC value of 1 is obtained; however, random distri-
butions cause an AUC value of 0.5. Virtual screening that
performs better than a random discrimination of actives and
decoys retrieve an AUC value between 0.5 and 1, whereas an
AUC value lower than 0.5 represents the unfavorable case of
a virtual screening method that has a higher probability to
assign the best scores to decoys than to actives [77,79].
(2) Overall Accuracy (ACC): describes the percentage of correctly
classified molecules by the screening protocol (Eq. (7)). Testing
compounds are assigned a binary score value of zero
(compound not captured) or one (compound captured)
[77,79,88].
ACC ¼TP þ TN
N
¼A
N$Se þ
?
1 ?A
N
?
$Sp(7)
where, N is the total number of compounds in the testing database,
A is the number of true actives in the testing database.
(3) Overall specificity (SPC): describes the percentage of discarded
inactives by the particular virtual screening workflow. Inactive
test compounds are assigned a binary score value of zero
(compound not captured) or one (compound captured)
regardless to their individual fit values [77,79,88].
(4) Overall true positive rate (TPR or overall sensitivity): describes
the fraction percentage of captured actives from the total
number of actives. Active test compounds are assigned a binary
score value of zero (compound not captured) or one
(compound captured) regardless to their individual fit values
[77,79,88].
(5) Overall false negative rate (FNR or overall percentage of dis-
carded actives): describes the fraction percentage of active
compounds discarded by the virtual screening method. Dis-
carded active test compounds are assigned a binary score value
of zero (compound not captured) or one (compound captured)
regardless to their individual fit values [77,79,88].
4.1.9. In silico screening for new PPARg activators
S1R3H1, S1R5H8and S3R2H2were employed as 3D search queries
to screen the national cancer institute (NCI) 3D structural database.
Virtual screening was performed employing the “Best Flexible
Database Search” option implemented within the CATALYST
module of Discovery Studio. The hits were filtered according to
Lipinski’s [80] and Veber’s [81] rules and the remaining hits were
Dði;jÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
?NumHBDi? NumHBDj
?2þ?NumHBAi? NumHBAj
?2þ?NPi? NPj
?2
q
(3)
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
14
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040

Page 15

combined together and fitted against the three pharmacophores
using the “best fit” option within CATALYST. The fit values together
with the relevant molecular descriptors of each hit were
substituted in QSAR Eq. (1). The highest-ranking molecules based
on QSAR predictions were acquired and tested in vitro.
4.2. In vitro experimental studies
NCI hits were dissolved in DMSO in serial dilutions starting from
ca. 40 mM. The amount of DMSO did not exceed 1% of the final
concentration in each well. DMSO solution (1% v/v) was used as
negative control, while rosiglitazone was used as positive control at
a concentration range of 0.1e1.0 mM. All runs were repeated in
triplicates.
Bioassay was performed in a similar way to previously reported
methods [89]. Briefly, HepG2 cells (ATCC, Manassas, USA) were
cultured in MEM with L-glutamine (Invitrogen Corporation, Carls-
bald) supplemented with non-essential amino acids and sodium
pyruvate (Sigma Aldrich, Germany), penicillin and streptomycin,
and 10% fetal bovine serum at 37?C in a humidified atmosphere
containing 5% CO2in air. One day prior to transfection, cells were
seeded into culture plates to reach about 60% confluence. Trans-
fection was carried out using lipofectin (Invetrogen Corporation,
Carlsbald) according to manufacturer’s recommendation. In brief,
the transfection reagent was mixed with optimum serum free
medium (Invitrogen Corporation, Carlsbald) and incubated for
30e45 min before adding to the plasmid cocktail and incubated
further for 15 min before adding another portion of serum free
media and then pipetted slowly on the pre-washed cells and
incubated overnight. The plasmids used in this study were PPREx3-
Tk-Luc (a kind gift by Professor Ronald M. Evans of Salk Institute for
Biological Studies, USA), pSV-sport PPARg2 and pSV-sport RXRa
both supplied by Professor Bruce Spiegelman of Harvard Medical
school through Addgene website (www.addgene.org) and pRKTK
plasmid (from Promega Corporation, USA).
Cells were transfected with 0.5 mg of the PPREx3-Tk-Luc, 0.17 mg
pRLTK, 0.5 mg pSV-sport PPARg2, and 0.5 mg pSV-sport RXRa plas-
mids. After overnight incubation, the transfected cells were treated
withtestedcompounds’andincubatedover24 hat37?Cand5%CO2.
Bioactivity readings were measured by luminescence quantifi-
cations using Dual-Glo Luciferase Assay System (Promega Corpo-
ration, USA) according to manufacturer’s protocol. Briefly, 70mL
Dual-Glo luciferase reagent was added to the media of transfected
cells and incubated for 10 min and luminescence signals were read
using Glomax?96 microplate luminometer. Following that, 70 mL
Stop and Glo reagent was added into the wells and incubated
further for 10 min and luminescence signal were read again using
the same protocol.
The fold changes of luciferase ratio caused by the tested hits
were calculated using the following formula:
Fold Changes of Luciferaseration
Luciferase RatioHIT? Luciferase RationNTC
Luciferase RatioDMSO? Luciferase RationNTC
where, NTC¼non-transfected cells. Statistical analysis were con-
ducted using Minitab 15 software and statistical significance were
determined using Analysis of Variance (ANOVA) tests with p<0.05
considered significant.
¼
Acknowledgements
This project was partly funded by Malaysian Ministry of
Science, Technology and Innovation, Grant Scheme 311/IFN/
69230111. The authors also thank the Deanship of Scientific
Research and Hamdi-Mango Centre for Scientific Research at the
University of Jordan for their generous funds.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at doi:10.1016/j.ejmech.2011.03.040.
References
[1] L.M. Hall, N. Sattar, J.M. Gill, Risk of metabolic and vascular disease in South
Asians: potential mechanisms for increased insulin resistance, Future Lip-
idology 3 (2008) 411e424.
[2] B.R. Johns, G.M. Reaven, PPAR-g agonists, insulin resistance and dyslipidemia:
not a simple relationship, Clinical Lipidology 5 (2010) 509e525.
[3] P. Arck, B. Toth, A. Pestka, U. Jeschke, Nuclear receptors of the peroxisome
proliferator-activated receptor (PPAR) family in gestational diabetes: from
animal models to clinical trials, Biology of Reproduction 83 (2010) 168e176.
[4] A. Hiukka, M. Maranghi, N. Matikainen, M.R. Taskinen, PPAR alpha: an
emerging therapeutic target in diabetic microvascular damage, Nature
Reviews Endocrinology 6 (2010) 454e463.
[5] A. Farce, N. Renault, P. Chavatte, Structural insight into PPAR g ligands
binding, Current Medicinal Chemistry 16 (2009) 1768e1789.
[6] R. Siersbaek, R. Nielsen, S. Mandrup, PPAR g in adipocyte differentiation and
metabolism e novel insights from genome-wide studies, FEBS Letters 584
(2010) 3242e3249.
[7] J.S. Bae, T.H. Kim, M.Y. Kim, J.M. Park, Y.H. Ahn, Transcriptional Regulation of
glucose sensors in pancreatic beta-cells and liver: an update, Sensors 10
(2010) 5031e5053.
[8] J. Bassaganya-Riera, R. Song, P.C. Roberts, R. Hontecillas, PPAR-g activation as
an anti-inflammatory therapy for respiratory virus infections, Viral Immu-
nology 23 (2010) 343e352.
[9] S.Z. Duan, M.G. Usher, R.M. Mortensen, PPARs: the vasculature, inflammation
and hypertension, Current Opinion in Nephrology and Hypertension 18
(2009) 128e133.
[10] B.D. Abbott, Review of the expression of peroxisome proliferator-activated
receptors alpha (PPARa), beta (PPARb), and gamma (PPARg) in rodent and
human development, Reproductive Toxicology 27 (2009) 246e257.
[11] E.R. Pearson, Pharmacogenetics and future strategies in treating hyper-
glycaemia in diabetes, Frontiers in Bioscience 14 (2009) 4348e4362.
[12] B.R. Henke, Peroxisome proliferator-activated receptor gamma (PPAR g)
ligands and their therapeutic utility, in: Progress in Medicinal Chemistry.
Elsevier, 2004, pp. 1e53.
[13] J. Choi, Y. Ko, H.S. Lee, Y.S. Park, Y. Yang, S. Yoon, Identification of (b-car-
boxyethyl)-rhodanine derivatives exhibiting peroxisome proliferator-acti-
vated receptor g activity, European Journal of Medicinal Chemistry 45 (2010)
193e202.
[14] J.H. Ahn, M.S. Shin, S.H. Jung, J.A. Kim, H.M. Kim, S.H. Kim, S.K. Kang, K.R. Kim,
S.D. Rhee, S.D. Park, J.M. Lee, J.H. Lee, H.G. Cheon, S.S. Kim, Synthesis and
structureeactivity relationship of novel indene N-oxide derivatives as potent
peroxisome proliferator activated receptor(PPARg) agonists, Bioorganic &
Medicinal Chemistry Letters 17 (2007) 5239e5244.
[15] Y. Lamotte, P. Martres, N. Faucher, A. Laroze, D. Grillot, N. Ancellin,
Y. Saintillan, V. Beneton, R.T. Gampe Jr., Synthesis and biological activities of
novel indole derivatives as potent and selective PPAR g modulators, Bio-
organic & Medicinal Chemistry Letters 20 (2010) 1399e1404.
[16] S.D. Debenham, A. Chan, F.W. Lau, W. Liu, H.B. Wood, K. Lemme, L. Colwell,
B. Habulihaz, T.E. Akiyama, M. Einstein, T.W. Doebber, N. Sharma, C.F. Wang,
M. Wu, J.P. Berger, P.T. Meinke, Highly functionalized 7-azaindoles as selective
PPARg modulators, Bioorganic & Medicinal Chemistry Letters 18 (2008)
4798e4801.
[17] X.-Y. Ye, S. Chen, H. Zhang, K.T. Locke, K. O’Malley, L. Zhang, R. Srivastava,
B. Miao, D. Meyers, H. Monshizadegan, D. Search, D. Grimm, R. Zhang, J. Lippy,
C. Twamley, J.K. Muckelbauer, C. Chang, Y. An, V. Hosagrahara, L. Zhang,
T.J. Yang, R. Mukherjee, P.T.W. Cheng, J.A. Tino, Synthesis and structure-activity
relationships of 2-aryl-4-oxazolylmethoxy benzylglycines and 2-aryl-4-thia-
zolylmethoxybenzylglycinesasnovel,potentPPARaselectiveactivators-PPAR
a and PPAR g selectivity modulation, Bioorganic & Medicinal Chemistry Letters
20 (2010) 2933e2937.
[18] R.C. Desai, W. Han, E.J. Metzger, J.P. Bergman, D.F. Gratale, K.L. MacNaul,
J.P. Berger, T.W. Doebber, K. Leung, D.E. Moller, J.V. Heck, S.P. Sahoo, 5-Aryl
thiazolidine-2,4-diones: discovery of PPAR dual a/g agonists as antidiabetic
agents, Bioorganic & Medicinal Chemistry Letters 13 (2003) 2795e2798.
[19] R.C. Desai, D.F. Gratale, W. Han, H. Koyama, E. Metzger, V.K. Lombardo,
K.L. MacNaul, T.W. Doebber, J.P. Berger, K. Leung, R. Franklin, D.E. Moller,
J.V. Heck, S.P. Sahoo, Aryloxazolidinediones: identification of potent orally
active PPAR dual a/g agonists, Bioorganic & Medicinal Chemistry Letters 13
(2003) 3541e3544.
[20] C. Santini, G.D. Berger, W. Han, R. Mosley, K. MacNaul, J. Berger, T. Doebber,
M. Wu, D.E. Moller, R.L. Tolman, S.P. Sahoo, Phenylacetic acid derivatives as
hPPAR agonists, Bioorganic & Medicinal Chemistry Letters 13 (2003)
1277e1280.
B.O. Al-Najjar et al. / European Journal of Medicinal Chemistry xxx (2011) 1e17
15
Please cite this article in press as: B.O. Al-Najjar, et al., Discovery of new nanomolar peroxisome proliferator-activated receptor g activators via
elaborate ligand-based modeling, European Journal of Medicinal Chemistry (2011), doi:10.1016/j.ejmech.2011.03.040