QSAR studies of paeonol analogues for inhibition of platelet aggregation

Article abstract of DOI:10.1016/j.bmc.2005.07.027

Various paeonol analogues were synthesized and tested in vitro as inhibitors of platelet aggregation. Structural properties (or descriptors) of paeonol analogues were calculated and the structure-activity relationships were determined. Several multiple linear and nonlinear regression models and back-propagation neural network model were tested and the latter using relative positive charge, hydration energy, and hydrophilic factor as inputs gave the best data fitting with R² = 0.89 and qpre2=0.66. The correlation coefficient between antiplatelet inhibition activity with an interaction energy between the paeonol compounds with COX-1 enzyme is only 0.39.

Full text of DOI:10.1016/j.bmc.2005.07.027

Bioorganic & Medicinal Chemistry 13 (2005) 5996–6001
QSAR studies of paeonol analogues for inhibition of
platelet aggregation
Mukesh Doble,a,* S. Karthikeyan,b P. A. Padmawarc and K. G. Akamanchic
aDepartment of Biotechnology, Indian Institute of Technology Madras, Chennai 600036, India
bCenter for Biotechnology, Anna University, Chennai 600025, India
cPharmaceutical Division, Department of Chemical Technology, University of Mumbai, Matunga 400019, India
Received 31 May 2005; revised 5 July 2005; accepted 6 July 2005
Available online 2 September 2005
Abstract—Various paeonol analogues were synthesized and tested in vitro as inhibitors of platelet aggregation. Structural properties
(or descriptors) of paeonol analogues were calculated and the structure–activity relationships were determined. Several multiple lin-
ear and nonlinear regression models and back-propagation neural network model were tested and the latter using relative positive
charge, hydration energy, and hydrophilic factor as inputs gave the best data fitting with R2 = 0.89 and qp2re ¼ 0.66. The correlation
coefficient between antiplatelet inhibition activity with an interaction energy between the paeonol compounds with COX-1 enzyme is
only 0.39.
Ó 2005 Elsevier Ltd. All rights reserved.
1. Introduction
activity. A quantitative structure–activity relationship
(QSAR) study was carried out on these compounds
and reported in this paper. To develop QSAR models
with good predictive capability, appropriate molecular
descriptors have to be chosen. Several softwares are
available to calculate almost 2000 descriptors of a mol-
ecule. The problem that is faced frequently by a
researcher is that of a small number of observations
and a large number of molecular parameters in the
descriptor pool, and one has to select the best descrip-
tors that represent the system from this large set. At
times, selecting the wrong set of descriptors could result
in chance correlations or missed understanding. Accord-
ing to researchers, quality of QSAR results mainly
depends on two factors, namely, the kind of molecular
descriptor and the method used to extract the useful
molecular information. These problems are addressed
by applying several statistical techniques, such as cluster
(dissimilarity) analysis and principal component analy-
sis. The technique of cluster analysis involves grouping
the data into subsets or clusters, such that those within
each cluster are more closely related to one another than
the data assigned to different clusters. The central theme
of cluster analysis is calculating the degree of dissimilar-
ity between the individual objects that are being
clustered. Principal component analysis involves a
mathematical procedure that transforms a number of
correlated variables into a smaller number of uncorre-
lated variables called principal components.
Blood platelets are involved in the early stages of many
cardiovascular diseases namely myocardial infarction,
diabetic vascular disorders, and arteriosclerosis. Aspirin
and dipyridamole had been used as antiplatelet drugs,
but had been identified with side effects. In addition to
sodium salicylate and 3,4-dihydroxybenzoic acid, paeo-
nol, an analgesic isolated from the root of Paeonia
suffritcosa Andr and Cynanchm paniculatum (Bunge) kit-
aq, coumarin, chromone, and bicyclic and tricyclic
derivatives of 4(3H)pyrimidnone1 and 1-phenylbenzimi-
dazoles2 have also been found to have antiplatelet and
antiischamic activity. A number (19 analogues) of ana-
logues of this 1-(2-hydroxy phenyl)ethanone were syn-
thetically prepared by substituting at the C-4 and the
C-5 positions of the aromatic ring and were tested in vi-
tro for antiaggregation activity triggered by ADP in
blood plasma. Akamanchi et al.3 reported the synthesis,
and activity studies of these paeonol compounds.
A systematic understanding of the various substituents
on the activity of the analogues helps one to rationalize
the findings and also design compounds with enhanced
Keywords: Platelet aggregation; Heat of formation; Hydrophilic factor;
Highest occupied molecular orbital; Neural network models.
*
Corresponding author. Tel.: +91 44 2257 4107; e-mail:
mukeshd@iitm.ac.in
0968-0896/$ - see front matter Ó 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.bmc.2005.07.027

M. Doble et al. / Bioorg. Med. Chem. 13 (2005) 5996–6001
5997
2. Results and discussion
Dendrogram
5
4
3
2
1
0
Table 1 gives the platelet aggregation inhibitory activity
of the 20 compounds. Compound 3 has the highest
inhibitory activity (65.36% vs 36.97% for aspirin and
36.31% for paeonol (compound 4) at 300 lM against
5 lM ADP). Figure 1 shows the cluster analysis of the
inhibition data. It appears that the data set could be
grouped into three sets based on their activity, the low
(0–15%), medium (15–35%), and high activity (35–
100%) sets. The high activity set has only one compound
(3) that has R1 = H and R2 = OH. The low and medium
sets have 9 and 10 compounds, respectively. There does
not seem to be any pattern between the substituent and
activity.
Table 2 lists only five descriptors, namely, relative posi-
tive charge (RPCG), hydration energy (HE), heat of for-
mation (HF), HOMO (highest occupied molecular
orbital), and hydrophilic factor (Hyf), that have a corre-
lation coefficient value greater than 0.5 with antiaggre-
gation activity. The remaining 414 descriptors have a
correlation coefficient less than 0.5 and hence are consid-
ered not significant. A negative correlation coefficient
indicates increasing the descriptor decreases the anti-
platelet aggregation activity. Cross-correlation studies
between these five descriptors indicate that RPCG is
highly correlated with HF (ꢀ0.9) and HE with Hyf
(ꢀ0.77). Thus, one could select either RPCG or HF
and HE or Hyf and perform regression analysis and
modeling studies with only three descriptors. Studies
carried out by Tavet et al.5 with 1,4 bis(poly and mono
O bjects
Figure 1. Results of cluster analysis.
Table 2. Descriptors that are highly correlated with antiplatelet
inhibition activity
Descriptors
Correlation
coefficient
Relative positive charge
Hydration energy (kcal/mol)
Heat of formation
HOMO
ꢁ0.557
ꢁ0.545
ꢁ0.542
0.512
Hydrophilic factors
0.645
methoxybenzoyl)-piperazines indicated that compounds
bearing hydrophilic or slightly hydrophobic substituents
exhibited good antiaggregation activity. Voskressensky
et al.6 have observed that logP value of 2.5 was optimal
lipophilicity for analogues of pyrrolopyridines to exhibit
antiplatelet activity. Lipophilicity was also observed as
an important structural parameter by De Marco et al.7
with derivatives of nipecotic acid anilides. Antiplatelet
activity of 2-substituted phenyl and benzimidazolyl-5-
methyl-4-(3-pyridyl) imidazoles was explained with
micelle-water partition coefficient, while CLOGP (a
descriptor used to estimate hydrophobicity of benzene
derivatives) failed to correlate with the activity.8
Table 1. Inhibition activity of paeonol analogues
O
R2
CH3
R1
OH
Compound
No.
R1
R2
% Inhibition
1
2
H
H
4.10
29.44
65.36
36.31
24.47
3.99
OH
H
H
OH
3
Cyclic peptides related to L-aspartic acid, N-[3-(amino-
methyl)benzoyl]-D-valyl-N2-methyl-L-arginylglycyl-,
cyclic (41 ! 1)-peptide (DMP 757) containing heterocy-
clic and other modified linking moieties exhibited in vitro
antiplatelet aggregation activity.9 The authors postulat-
ed that the lipophilic binding pocket favored a similar
linking moiety for high activity. Our studies also indicate
higher the hydrophilic factor the higher the antiaggrega-
tion activity. Also, for our system log P has a correlation
coefficient of only ꢁ0.36, and hence has not been includ-
ed in model development. Studies with anilides and
phenyl esters of piperidine-3-carboxylic acid have indi-
cated that high lipophilicity and increased electron den-
sity of the phenyl ring increased the activity.10 Similar
observations were made with bis-nipecotamides with
varying aromatic bridges connecting the two nipecota-
4
OCH3
H
H
OCH3
5
6
C
H
Cl
7
H
Br
8.20
6.87
8
H
Br
9
H
8.78
1.49
10
11
12
13
14
15
16
17
18
19
20
F
H
H
CH3
F
5.44
H
CH3
10.00
3.00
0.20
H
NO2
H
NO2
H
NHCOCH3
30.40
15.84
20.86
24.43
17.60
1.00
H
NHCOCH3
Cl
H
CH3
Cl
F
NO2
CH3

5998
M. Doble et al. / Bioorg. Med. Chem. 13 (2005) 5996–6001
mide moieties. Shen et al.2 have successfully used an elec-
tronic descriptor to describe the activities of 1-phenyl-
benzimidazoles. Our studies also show a positive
correlation between the HOMO and activity. In addi-
tion, our studies show a negative correlation between
heat of formation and activity, that is, more negative
the heat of formation value (more thermodynamically
stable is the molecule) higher the antiplatelet aggregation
activity.
ponents for all the data. It is clearly seen that there are
two distinct groups; the lower group relates to the medi-
um and high activity compounds, and the upper group
relates to the low activity data set. Of course, there are
one or two compounds from the other group present
in the wrong cluster.
Table 3 lists the various linear and nonlinear multiple
regression models attempted to fit the data with the cor-
responding statistics. The table also lists the results of
modeling studies carried out with an artificial neural net-
work with three inputs, one hidden layer with three pro-
cessing elements using TanH transfer function and Delta
learning rule. Excellent model fitting is obtained with
artificial neural network (ANN) models (best results
are R2 of 0.89, Ra2dj of 0.74, qp2re of 0.68, and F of
145.5 for a 3-3-1 model with relative positive charge,
hydration energy, and hydrophilic factor as input vari-
ables). Figure 3 compares the ANN model predictions
with the data. The cross-validated q2 is good, indicating
that the model has good predictive capabilities.
2.1. Modeling
A generic QSAR model has been suggested by Wang
et al.11 as,
log½Activityꢂ ¼ A log½penetrationꢂ þ B log½interactionꢂ
þ C;
where penetration could be considered as a hydropho-
bicity controlled process. Interaction of the molecule
and target could be due to electronic or steric12 or elec-
trophilic in nature or due to the molecular orbital con-
tent of the molecule.
To test further the predictive capability of the model and
also determine if the ANN was overtrained, the data
were divided into two sets, namely, a learning set con-
taining 14 data and a testing/validation set containing
6 data. The learning of the ANN was carried out using
the former and the prediction with the latter data set.
A 3-3-1 back propagation model using TanH transfer
function and Delta learning rule gave a R2 of 0.87 and
Principal component analysis is an unsupervised learn-
ing algorithm and it is estimated from only the five
descriptors short-listed and not using all the 419 descrip-
tors. Three and four principal components are needed to
explain 95.35% and 99.2% of the observed variance,
respectively. Figure 2 plots the first three principal com-
70
60
50
40
30
20
10
0
High
Medium
Low
0
20
40
60
80
Inhibition % (Data)
Figure 3. Comparison of back-propagation neural network model
(3-3-1, inputs: HF, HOMO, Hyf, Delta learning rule, and TanH
transfer function) prediction with data.
Figure 2. Plot of first three principal components of all the paeonol
analogues.
Table 3. Regression equations relating the descriptors with antiplatelet inhibition activity
Regression equation
R2
R2adj
q2pre
RMSSE
F
18.529 + 40.610 Hyf
*
0.416
0.290
0.260
0.577
0.577
0.380
0.260
0.240
0.498
0.498
0.270
0.200
0.180
0.267
0.267
11.600
12.900
13.200
9.950
12.820
7.350
6.320
ꢁ17.255 ꢁ 0.414 HF
*
172.700 + 16.592 HOMO
*
3.962 + 0.007 HF ꢁ 1.710 HOMO + 47.240 Hyf
24.550
24.550
*
*
*
4.209 + 1.492 HF ꢁ 2.456 HOMO ꢁ 18.170 Hyf + 0.009
9.950
*
*
*
*
HF2 + 0.425 HOMO2 + 126. 580 Hyf2
Back-propagation ANN 3-3-1, inputs: HF, HOMO, Hyf, TanH, Delta
Back-propagation ANN 3-3-1, inputs: RPCG, HE, Hyf, TanH, Delta
*
*
0.780
0.890
0.700
0.740
0.640
0.680
7.160
4.980
63.800
145.500

M. Doble et al. / Bioorg. Med. Chem. 13 (2005) 5996–6001
5999
80
60
40
20
0
40
30
20
10
0
10
20
30
40
0
60
7
0
Inhibition % (Data)
0
10
20
30
40
Figure 4. Comparison of back-propagation neural network model (3-2-1, inputs: RPCG, HE, Hyf, Delta learning rule, and TanH transfer function)
prediction with learning data set (n = 14). Inset prediction with testing/validation data set (n = 6). Bars represent 1 RMSSE.
0.43 for the training and prediction sets, respectively. A
3-2-1 model gave a R2 of 0.85 ðRa2dj of 0.73Þ for the
60
training set and a R2 of 0.57 for the prediction set. A
3-1-1 model gave a R2 of 0.74 and 0.28 for the training
and prediction sets, respectively. This study indicated
that a 3-2-1 ANN model was the best one with respect
to the training, as well as prediction R2. Figure 4 com-
pares the ANN model predictions with the learning, as
well as the test data sets for the 3-2-1 network.
40
20
0
2.2. Binding of paeonol analogues to COX-1
-55
-45
-35
-25
Aspirin exerts its anti-inflammatory effects through
selective acetylation of serine 530 on prostaglandin H2
synthase (PGHs) (COX-1).13 The mechanism of aspi-
rinÕs antiplatelet activity is thought to work predomi-
nantly through its inhibition of cyclo-oxygenase.14 The
twelve amino acids surrounding the active sites have
been identified and they are 116 VAL, 120 ARG, 349
VAL, 352 LEU, 353 SER, 355 TYR, 359 LEU, 518
PHE, 523 ILE, 527 ALA, 530 SER, and 531 LEU.15
Molecular mechanics (MM) interaction energy between
these amino acids and various paeonol analogues was
determined individually. Figure 5 shows a plot of %
inhibition activity as a function of the binding energy
for the paeonol analogues, where the binding energy
or interaction energy is the difference between the total
energy of the complex and the energy of the correspond-
ing paeonol. As seen from the figure, as the interaction
energy increases the antiplatelet inhibition activity possi-
bly also increases, although such a conclusion cannot be
made with certainty since the correlation coefficient is
only 0.39.
Energy Difference (Docked-Inhibitor) Energy (kcals/mol)
Figure 5. Relation between % inhibition activity versus binding energy
for paeonol analogues with COX-1.
nent analysis clearly shows two data clusters, represent-
ing the lower and the higher activity sets. Several
multiple linear and nonlinear regression models and
back propagation neural network model were developed
and the latter using relative positive charge, hydration
energy, and hydrophilic factor as inputs gave the best
data fitting, as well as predictive R2. The descriptors
shortlisted in our study are the same as the descriptors
mentioned by other researchers for several other series
of anti-aggregation compounds, which indicates, irre-
spective of the compound, the mechanism of action is al-
most same. This QSAR approach developed using
structural descriptors can help in understanding the
properties that contribute to the action of a drug, there-
by designing drugs with higher activity. Overfitting of
the data will lead to a model with poor predictive capa-
bility. Hence a balance needs to be struck between over-
fitting the data and good predicative capability. This
point was clearly seen when the data were divided into
two groups, one data set used for learning/training the
network and the other used for only validation/testing.
The number of processing elements in the hidden layer
was changed until the R2 for the learning and the valida-
tion/testing sets were highest. Such an exercise gave a 3-
2-1 back-propagation network, which had good learning
as well as predictive capability.
3. Conclusions
Twenty paeonol analogues were synthesized and tested
for antiplatelet aggregation activity in vitro. QSAR
studies have indicated that five best descriptors, namely,
relative positive charge, hydration energy, heat of
formation, HOMO, and hydrophilic factor can explain
the observed anti-aggregation activity. Principal compo-

6000
M. Doble et al. / Bioorg. Med. Chem. 13 (2005) 5996–6001
4. Experimental
and the correlation coefficients for all of them with inhi-
bition activity were determined.
4.1. Synthesis
4.3. Statistical methods
More details of the synthesis of these compounds are
mentioned in Akamanchi et al.3 Although many syn-
thetic methods can be used for the preparation of
these compounds, the Fries rearrangement4 was pre-
ferred since the reactions could be performed easily
and the regioisomers formed could be readily separat-
ed with good yield. Compound 2 was prepared by the
Friedel–Crafts reaction of resorcinol with glacial acetic
acid in the presence of anhydrous zinc chloride. Selec-
tive methylation of compound 2 using excess methyl
iodide and lithium carbonate in dimethyl formamide
gave paeonol (4). The Fries rearrangement of 4-meth-
oxyphenyl acetate gave predominantly compound 3.
Selective methylation of compound 3 with methyl
iodide and potassium carbonate in acetone gave 5.
Compounds 6–20 were synthesized by the Fries rear-
rangement by heating a mixture of anhydrous alumi-
num chloride and corresponding substituted phenyl
acetate. The synthesized compounds were character-
ized by their melting point (Campbell melting point
apparatus), IR spectra (Shimadzu IR 408 using KBr
disk) and 1H NMR spectra (Varian EM 360,
60 MHz).
Statistical techniques, such as Cluster Analysis, Princi-
pal Component Analysis, and Nonlinear regression
analysis, were performed on the data set using SY-
STATÒ software (SPSS Inc., USA). The same software
was used for estimating dissimilarity distance and cor-
relation coefficients between various molecular parame-
ters. Neural Network simulations were carried out
using NEURALWAREÒ software (NeuralWare, Inc.,
PA, USA). Back-propagation neural network with
TanH transfer function was used in all cases with Del-
ta-learning rule and 16 epochs. Estimation of dissimi-
larity distance and correlation coefficient was carried
out to identify the best set of molecular descriptors
to be used in the regression and neural network
models.
The goodness of the regression fits were estimated
using parameters, such as R2, Ra2dj, qp2re (also known
as validation R2), RMSSE, and F ratio, which are de-
fined below,
R2 ¼ 1 ꢁ SSE=TSS
Platelet aggregation inhibitory activity of these com-
pounds was studied in vitro at King Edward VII
Memorial Hospital, Mumbai, India, from blood sam-
ples taken from donors. Platelet aggregation inhibito-
R2adj ¼ 1 ꢁ ðn ꢁ 1Þð1 ꢁ R2Þ=ðn ꢁ p ꢁ 1Þ
q2pre ¼ 1 ꢁ PRESS=TSS
ry activity was measured using
a
four-channel
aggrecorder3 (more details are given in Akamanchi
et al., 1999). The concentration of adenosine diphos-
phate eliciting the full biphasic aggregation response
(ꢀ5 lM) was established using platelet-rich plasma
from each donor before the determination of inhibito-
ry potency. Test drugs were prepared by dissolving
them in methanol to prepare 2% solution. Maximum
percent aggregation (MPA) obtained from the aggre-
gation trace of control (methanol) and treated (test
compound) was used to calculate % inhibition with
the following formula,
RMSSE ¼ mean sum of square of the error
pffiffiffiffiffiffiffiffiffiffiffiffiffi
¼
SSE=n
F ¼ ðn ꢁ 2ÞR2=ð1 ꢁ R2Þ
where n, number of data points; p, number of
parameters.
X
2
TSS ¼
ðydata;i ꢁ yavgÞ
% Inhibition ¼ 100 ꢃ ð1 ꢁ MPA of test compound=
MPA of controlÞ:
X
2
SSE ¼
yavg
ðymodel;i ꢁ ydata;i
Þ
X
¼
ydata;i=n
4.2. Molecular modeling studies
The molecule structures of the compounds were built
using HYPERCHEMÒ (HYPERCUBE Inc., USA)
and their minimum energy conformation was deter-
mined first by minimizing the structures using molecular
mechanics MM+ force field, followed by minimizing the
structures with semi-empirical quantum mechanics
(AM1 method with restricted Hartree–Fock (RHF))
and finally with an ab initio method (STO-3G minimal
basis set). The constitutional, topological, geometrical,
charge, and quantum mechanical descriptors (419 in to-
tal) were estimated using the software DRAGONÒ
(Milano Chemometerics, Italy), and HYPERCHEMÒ
ydata, i = data points; ymodel, i = model predictions.
PRESS = A model is developed with (n ꢁ 1) data
points and the nth point is predicted using this model.
This exercise is carried out for all the points. The sum
of squares of the difference between these predicted
data (using the Ôleave-one-outÕ scheme) and the actual
values is called PRESS. Predictive capability of a
model is expressed by q2pre. RMSSE is an indication
of the mean deviation of the prediction from the data.
A large F indicates the model fit is not a chance
occurrence.

M. Doble et al. / Bioorg. Med. Chem. 13 (2005) 5996–6001
6001
8. Tanaka, A.; Nakamura, K.; Nakanishi, I.; Fujiwara, H. J.
Med. Chem. 1994, 37, 4563.
References and notes
9. Wityak, J.; Fevig, J. M.; Jackson, S. A.; Johnson, A. L.;
Mousa, S. A.; Parthasarathy, A.; Wells, G. J.; Degrado, W.
F.; Wexler, R. R. Bioorg. Med. Chem. Lett. 1995, 5,
2097.
10. De Candia, M.; Summo, L.; Carrieri, A.; Altomare, C.;
Nardecchia, A.; Cellamare, S.; Carotti, A. Bioorg. Med.
Chem. 2003, 11, 1439.
11. Wang, X.; Yin, C.; Wang, L. Chemosphere 2002, 46,
1045.
12. Schultz, T. W.; Cronin, M. T. D.; Walker, J. D.; Aptula,
A. O. J. Mol. Struct. (THEOCHEM) 2003, 622, 1.
13. Loll, P. J.; Picot, D.; Garavito, R. M. Nat. Struct. Biol.
1995, 2, 605.
1. Roma, G.; Braccio, M. D.; Carrieri, A.; Grossi, G.;
Leoncini, G.; Signorello, M. G.; Carotti, A. Bioorg. Med.
Chem. 2003, 11, 123.
2. Shen, Q.; Lu, Q. Z.; Jiang, J. H.; Shen, G. L.; Yu, R. Q.
Eur. J. Pharm. Sci. 2003, 20, 63–71.
3. Akamanchi, K. G.; Padmawar, P. A.; Thatte, U. M.;
Rege, N. N.; Dahanukar, S. A. Pharm. Pharmacol.
Commun. 1999, 5, 323.
4. Martin, R. Org. Prep. Proced. Int. 1992, 24, 373.
5. Tavet, F.; Lamouri, A.; Heymans, F.; Dive, G.; Touboul,
E.; Blavet, N.; Godfroid, J. J. J. Lipid Mediators Cell
Signalling 1996, 15, 145.
6. Voskressensky, L. G.; De Candia, M.; Carotti, A.;
Borisova, T. N.; Kulikova, L. N.; Varlamov, A. V.;
Altomare, C. J. Pharm. Pharmacol. 2003, 55, 323.
7. De Marco, M.; De Candia, M.; Carotti, C. S.; De Candia,
E.; Altomare, C. J. Pharm. Sci. 2004, 22, 153.
14. Bradford, B.; Worrall, M. D.; Karen, C. J. Neurosurgeons
Focus 2000, 8, 14.
15. http://bioportal.weizmann.ac.il/oca-bin/vcofc.cgi?num=
8&pdb_id=1pth&xid=00204200001115748120.