Multiple Field Three Dimensional Quantitative

StructureActivity Relationship (MF-3D-QSAR)

QI-SHI DU,1,2,3 RI-BO HUANG,1,4 YU-TUO WEI,1 LI-QIN DU,1 KUO-CHEN CHOU3
1

Guangxi University, Guangxi Key Laboratory of Subtropical Bioresource Conservation and

Utilization, Nanning, Guangxi, 530004, China
2
Hainan Normal University, Department of Chemistry, Haikou, Hainan, 571158, China
3
Gordon Life Science Institute, San Diego, California 92130, USA
4
Guangxi Academy of Sciences, 98 Daling Road, Nanning, Guangxi 530004, China
Received 15 February 2007; Revised 5 April 2007; Accepted 17 April 2007
DOI 10.1002/jcc.20776
Published online 8 June 2007 in Wiley InterScience (www.interscience.wiley.com).

Abstract: A new drug design method, the multiple eld three-dimensional quantitative structureactivity relationship (MF-3D-QSAR), is proposed. It is a combination and development of classical 2D-QSAR and traditional 3DQSAR. In addition to the electrostatic and van der Waals potentials, more potential elds (such as lipophilic potential, hydrogen bonding potential, and nonthermodynamic factors) are integrated in the MF-3D-QSAR. Meanwhile, a
principal component analysis (PCA) and iterative double least square (IDLS) technique is developed for predicting
the bioactivity of query drug candidates. As an example, the MF-3D-QSAR is applied to the design of neuraminidase inhibitor and to prove its predictive power, and some useful ndings are obtained for developing drugs against
inuenza virus.
q 2007 Wiley Periodicals, Inc.

J Comput Chem 29: 211219, 2008

Key words: 3D-QSAR; potential elds; drug design; neuraminidase inhibitor; H5N1 inuenza

Introduction
The MF-3D-QSAR is actually a direct extension and development of the traditional 2D-QSAR. Typically, the 2D-QSAR
approach is represented by a linear free energy equation13
logAb ab fhydr Xh ae felec Xe ast fst Xs

ysis),8,9 the interaction free energy ui(rj) of sample i at space

points rj is correlated to the bioactivities of molecule samples
through a linear equation,
N
X

bj ui rj pKi

(2)

j1

(1)

where Ab is related to either the receptor binding afnity or a

specic biological activity. Each of the terms in eq. (1) refers
to the property of a drug candidate that can affect either receptor recognition or activation. Each term is a linear or quadratic
function of a corresponding physicochemical parameter X. Commonly parameters are: (i) the n-octanolwater partition coefcient for a hydrophobic term, (ii) Taft Es parameter for steric
effect, (iii) Hammett constants to describe electronic effects,
and (iv) the molar refractivity to account for dispersion forces.
Typically, hydrophobic ( fhydr), electronic ( felec), and steric ( fst)
parameters of the ligands are used. Lately, some nonthermodynamic properties were included in the list, such as topology
indices.47
The 3D-QSAR approach builds a 3D model of the receptor
cavity, where a cubic grid is established by a set of space points
{rj}. Represented by CoMFA (comparative molecular eld anal-

where pKi refers to the logarithm bioactivity of sample i, which

is proportional to the binding free energy between ligand and receptor. In eq. (2) the {bj} is the sensitive coefcients of space
points. Usually the interaction energy contribution ui(rj) of sample i at point rj is evaluated by the following equation,8,9
ui rj fi;jelec fi:jvdw

Mi
X
qk

r
k1 j;k

Mi
X
k1

"

ek

k
rj;k

12

6 #

k
2
(3)
rj;k

In eq. (3) only the electrostatic interaction and van der Waals
interaction are calculated using Coulomb equation and Lennard-

Correspondence to: Q.-S. Du; e-mail: duqishi@yahoo.com or K.-C.

Chou; e-mail: kcchou@gordonlifescience.org

q 2007 Wiley Periodicals, Inc.

Du et al. Vol. 29, No. 2 Journal of Computational Chemistry

212

Jones 6-12 equation, respectively. In comparison with the linear

free energy equation (1) of the 2D-QSAR, the interaction energy
equation (3) of the 3D-QSAR may be over simplied. Many
useful factors, such as lipophilicity,1016 which may have important contribution to the bioactivity, are lost. Great efforts have
been made to include molecular lipophilicity potential and other
molecular properties in the 3D drug design strategy.8,1722
Except for the molecular lipophilicity eld, more molecular
potential elds are included in CoMFA, such as the H-bonding
eld,23 desolvation eld, and molecular orbital eld (HOMO
and LUMO).19 However, it is very difcult to integrate these
potential elds into the 3D-QSAR because of some technique
problems. At this point the 3D-QASAR is a setback from the
traditional 2D-QSAR. In CoMSIA,17,18 a different type of 3DQSAR, similarity indices are calculated instead of interaction
energies. Each molecule from the training set is compared to a
and the charge, hydrophocommon probe with a radius of 1 A
bicity, and hydrogen-bond property equal to 1. The functional
form of elds is selected to be Gaussian function with an attenuation factor  0.3.17,18 Compared to the Lennard-Jones and
Coulomb potentials, the Gaussian-type function has the advantage of using all grid points inside and outside the molecules,
and no arbitrary cutoffs are required. Five different similarity
elds are calculated: steric, electrostatic, hydrophobic, hydrogenbond donor, and hydrogen-bond acceptor. These elds cover the
major contributions to ligandprotein binding.17,18
In this study we try to solve the problem in a different way,
i.e., using MF-3D-QSAR by converting the molecular properties
in the classical 2D-QSAR to a 3D form, and introducing more
potential elds in the 3D-QSAR.

Theory and Method

The MF-3D-QSAR is directly developed from the traditional
2D-QSAR, in which all physical properties assume the 3D form.
The theoretical model of MF-3D-QSAR is schematically illustrated in Figure 1.
The MF-3D-QSAR has a 3D data matrix, DMNK, in which
the subscripts M, N, and K refer to the number of samples in
training set, the number of space points, and the number of
potential elds, respectively. The elements di,j,k of matrix DMNK
are the values of potential eld k at position rj of sample i.
Although the space points XN are described as one dimension in
the 3D data matrix DMNK of MF-3D-QSAR, actually, the space
points span a 3D Cartesian space, just like in traditional 3DQSAR.
Because the number of space points is much more than the
number of potential elds and the number of molecule samples,
i.e., N  K and N  M, over correlation problem may be
caused by it. For better predictive results, we have to reduce the
dimensions of space point. The principal component analysis
(PCA) method is used to reduce the original data matrix DMNK
to a smaller data matrix EMLK by performing the following calculations for every molecule sample i,
i

ZNN DiNK DtiNK

i 1; 2; . . . ; M

(4)

Figure 1. The three dimensional data structure and iterative solution

procedure of MF-3D-QSAR. The notation XN refers for space points,
SM for molecule samples in training set, FK for potential elds, and
AM for bioactivities of samples, In the three dimensional data matrix
DMNK, the space points N are much larger than the numbers of both
samples M and elds K, (N  M and K). The dimensions of space
points are reduced by using PCA, leading to a smaller data matrix
EMLK (L  N). Then the coefcients {ak} of potential elds, and
{bj} of the coefcients of principal components of space points are
determined by using IDLS procedure.
i

i i

ZNN PNN
N PNN

i 1; 2; . . . ; M

(5)
i

Equation (5) is an eigenvalue equation of matrix ZNN , PNN are

the eigenvectors, and LN are the eigenvalues. The L eigenvectors
i
PNL (L  N) with larger eigenvalues are used to take principal
components from DiNK,
it

EiLK PNL DiNK

i 1; 2; . . . ; M

(6)

In this way the original large data matrix DMNK is reduced to a

smaller data matrix EMLK (L  N). The PCA method is schematically illustrated in Figure 1.
On the basis of the above-mentioned theoretical model, the
foundational equation of MF-3D-QSAR is set as follows,
L
X
j1

bj

K
X

!
ak ei;j;k

pKi

i 1; 2; . . . ; M

(7)

k1

where coefcient set {ak} is for the potential elds, coefcient

set {bj} for the principal components of space points and ei;j;k is

Journal of Computational Chemistry

DOI 10.1002/jcc

Multiple Field 3D-QSAR

the element of the matrix EMLK. An iterative double least square

(IDLS) technique was developed to determine the values of the
coefcient sets {ak} and {bj} alternately by solving the 3D simultaneous linear equation set eq. (7). By using a set of initial
values of coefcients {a(0)
k }, the 3D data matrix EMLK is reduced
to a 2D data matrix GML with the elements given by
gi;j

K
X

ei;j;k ak

213

mental bioactivities and the predicted bioactivities, i.e.,

v
u
M 

u1 X
 n1


n1 2
Q
 Qn  t
pKi  pKi
 M i1
v
u M 
 
u1 X
n 2 
pKi  pKi
t
e
M i1


16

(8)

k1

Thus, the set of 3D simultaneous linear equation set [eq. (7)] is

reduced to a set of 2D equations, i.e.,
GML bL pKM

(9)

The above equation can be solved by using the least square

approach,
yielding the rst solutions of the sensitive coefcients
n o
1

bj

,
1

bL GtML GML 1 GtML pKM

(10)

Then the values of fbj g are used to reduce the 3D data matrix
EMLK to a 2D data matrix HMK with the elements given by
hi;k

K
X

ei;j;k bj

(11)

j1

Similarly, the set of 3D simultaneous linear equations (7) is

reduced to a set of 2D equations by eq. (11), as given by
1

HMK ak pKM

(12)

The above equation can be solved by using the least square ap1
proach, leading to the solution of sensitive coefcients fal g,
1

t
t
HMK 1 HMK
pKM
aK HMK

(13)

Then the values of {ak } are used for the new solutions of the
2
sensitive coefcients {bj } of the principal components of space
points. The above procedure is performed iteratively for n steps,
n
i.e., until reach the converged solutions as denoted by fak g and
n
n
n
{bj }. Now, the values of {ak } and {bj } can be used to predict the bioactivities of the ith drug candidate through the following equations:
it

EiLK PNL DiNK

pKipred

K
X
k1

n
ak

L
X

(14)
!

n
ei;k;j bj

(15)

j1

The convergence criterion for the iterative procedure is set as

the residue Q of the predicted bioactivity, which is the square
root of the summation of squared differences between experi-

Calculations and Results

As an example, the MF-3D-QSAR is applied to the study of the
structurebioactivity relationship for the neuraminidase (NA) inhibitor design against inuenza virus. The demonstrative example may be helpful for the new drug design against the H5N1
avian inuenza. Total 32 molecule samples are used in the training set. The molecular structures and bioactivities pIC50 of the
32 molecules24 are listed in Table1.
According to the standard 3D-QSAR methods, such as CoMFA
and CoMSIA, a cubic grid is set up surrounding the aligned ligand
molecules and thousands upon thousands potential values at these
grid points are calculated. However, not all these spatial points are
equally important, and the huge number of grid points may cause
the problem of over correlation. In the following demonstrative
example for MF-3D-QSAR, we are to use a different approach to
construct the 3D space surrounding the ligand molecules.
The X-ray crystal structure of N9 NA of inuenza virus is used
as the target protein. Its experimental structure (1f8b.pdb)25 is
shown in Figure 3, which is a complex with ligand DANA. Several
commercial available drugs against inuenza are designed based on
this structure.2629 The conformations, the locations in protein complex, and the coordinates of the 32 molecules are determined by
aligning them with the template DANA. The 11 functional residues
(Arg118, Glu119, Asp151, Arg152, Trp178, Ile222, Gln227,
Glu276, Arg292, Arg371, and Tyr406) and 7 framework residues27
(Arg156, Asn198, Arg224, Asn294, Glu425, Ala246, Thr247) are
taken from the NA protein. The functional residues interact directly
with the DANA and play important roles in the bioactivity. The
framework residues make no direct interaction with the DANA substrate. However, the framework residues hold the substrate and create a suitable environment for catalytic interaction.30,31 Total 279
atoms are included in the 18 residues.
For the current study, we need not to calculate the values of
potential elds at the points in a cubic grid surrounding the
aligned molecules as treated by other 3D-QSAR methods. Our
rationale is that, owing to the important roles of the functional
and framework residues on the bioactivity of a protein, the values of potential elds of the ligands at the positions of the 279
atoms in the framework and functional residues are more important than those of the other spatial points.
Here the 3D space surrounding the aligned ligand molecules
is constructed based on the crystal structure of NA receptor. In
contrast to that, the conventional 3D-QSAR does not require the
3D structure of a receptor, and hence has its advantage for the
case when the molecular structures of receptors are not available. However, when the structures of the receptors are available, it will make the prediction model more reasonable so as to

Journal of Computational Chemistry

DOI 10.1002/jcc

214

Du et al. Vol. 29, No. 2 Journal of Computational Chemistry

a

Table 1. The Structures and Experimental pIC50 of the 32 Molecules Used in the Benchmark Dataset.

M01 (pIC50 6.65)

M02 (pIC50 7.22)

M03 (9.52)

M04 (7.92)

M05 (7.66)

M06 (9.00)

M07 (9.00)

M08 (6.21)

M09 (6.28)

M10 (8.00)

M11 (9.00)

M12 (9.00)

(continued)

Journal of Computational Chemistry

DOI 10.1002/jcc

Multiple Field 3D-QSAR

215

Table 1. (Continued)

M13 (8.52)

M14 (6.22)

M15 (6.74)

M16 (6.68)

M17 (6.70)

M18 (6.57)

M19 (5.20)

M20 (8.05)

M21 (9.00)

M22 (5.43)

M23 (7.22)

M24 (5.70)

(continued)

Journal of Computational Chemistry

DOI 10.1002/jcc

Du et al. Vol. 29, No. 2 Journal of Computational Chemistry

216

Table 1. (Continued)

M25 (6.89)

M26 (6.74)

M27 (6.52)

M28 (6.70)

M29 (7.05)

M30 (6.28)

M31 (5.66)

M32 (5.70)

From Maring et al.24

provide much more accurate predicted results by constructing

the 3D potential elds based on the structural knowledge of the
protein receptors according to the procedure provided in this
paper. The locations of the functional residues and the framework residues in NA are shown in Fig. 2.
The values of ve potential elds produced by the 32 ligands at
the positions of the 279 atoms of functional and framework residues are calculated. The electrostatic potentials are computed using
classical Coulomb equation, the van der Waals potentials using 612 Lennard-Jones equation, the hydrogen bond donor and acceptor
potentials using 4-8 Lennard-Jones equation, and the lipophilic
potential using the equation suggested by Viswanadhan et al.:16
filip rj

Mi
X

lk
r
1
j;k
k1

(17)

where the Mi is the total number of atoms in molecule i, rj,k is the

distance between atom k and point rj, and lk is the lipophilic parameter16 of atom k in molecule i.
The IDLS technique described in Theory and Method section
is used in the structureactivity relationship study for the NA inhibitor design for inuenza virus based on the experimental data
listed in Table 1. The initial values of sensitive coefcients
0
of ve potential elds fak g are assigned to be 1. This is a
0
reasonable initial guess for fak g that means the roles of all
potential elds are equal. The curves of correlation coefcients
R vs iterations are shown in Figure 3(A), where Ra is for the
n
iterated coefcients fak g and Rb is for the iterated coefcients
n
fbj g. The residue Q between the predicted bioactivities and the
experimental bioactivities of molecules are shown in Figure
n
n
3(B), where Qa is for fak g and Qb is for fbj g. It has been
observed that, after 20 iterations, the iterative result is quite

Journal of Computational Chemistry

DOI 10.1002/jcc

Multiple Field 3D-QSAR

Figure 2. Active sites in NA (N9) of inuenza virus type A. The

residues (Arg118, Glu119, Asp151, Arg152, Trp178, Ile222,
Glu227, Glu276, Arg292, Arg371, and Tyr406), which are shown in
green, are dened as functional residues according to their contacts
with the DANA (rendered by element color). The residues (Arg156,
Asp198 (Asn198 in N9), Arg224, Asn294, and Glu425), which are
shown in red, are part of framework residues in the sense that they
make no direct contact with the DANA substrate, which interact
with the functional residues to hold them in place for binding and
catalysis with the substrate. The ligand DANA is in the center of
active cavity and rendered by element color. See ref. [27] for further
explanation of the gure. [Color gure can be viewed in the online
issue, which is available at www.interscience. wiley.com.]

smoothly converged. For the training dataset the correlation

coefcient is R 0.9698 and the average residue of calculated
bioactivity is Q 60.2916, when 21 principal components are
n
used. The converged sensitive coefcient set {ak } are listed in
Table2. In the Figure 3, we observed that the correlation coefcient R increases from 0.7895 to 0.9698, and the predicted residue Q decreases from 0.7534 to 0.2916 with the iterations. The
big improvement is from the better theoretical model and
advanced mathematical technique of MF-3D-QSAR.
n
n
With the two sets of coefcients fak g and fbj g, eqs. (14)
and (15) can then be used for the bioactivity prediction of query
ligands. Jackknife test32,33 is performed for evaluating the predictive ability of MF-3D-QSAR approach. The predicted bioactivities of the 32 molecules are listed in Table3 with correlation
coefcient Rpred 0.8194 and predictive residue Qpred
60.8578. In the jackknife test, we found that the predicted bioactivities of molecules change with the iterations. In the rst
several iterations the predicted activities uctuate remarkably.
Then the predicted activities converge to the values close to the
experimental bioactivities, as shown in Figure 4.

Conclusion
The MF-3D-QSAR is a direct extension of the classical 2D-QSAR
and retains the advantages of traditional 3D-QSAR. The MF-3D-

217

Figure 3. (A) The correlation coefcients of the calculated bioactiv(n)

ities in training set. Ra is for {a(n)
k } iteration and Rb is for {bj } iteration. The correlation coefcient R increases from 0.7895 to 0.9698.
(B) The residue Q between predicted bioactivities and experimental
bioactivities of molecules in training set. Qa is for {ak(n)} iteration
and Qb is for {b(n)
j } iteration. The residue Q decreases from 0.7534
to 0.2916 with the iterations. Promising improvements are from the
better theoretical model and the advanced mathematical technique
of MF-3D-QSAR. [Color gure can be viewed in the online issue,
which is available at www.interscience.wiley.com.]

QSAR makes it possible to integrate more potential elds

(including the non thermodynamic elds) in the 3D-QSAR
approach. Because more potential elds and the two sets of adjustable coefcients {ak} and {bj} are used, MF-3D-QSAR enhances
Table 2. The Initial Coefcients

Coefcients

20
fak g

fak g and the Converged

of the Five Potential Fields.

Potential
aelec
avdw
alip
ahbd
ahba

Initial

Converged

1.0
1.0
1.0
1.0
1.0

1.5365
8.6323
0.4440
2.2793
0.8046

ahbd, hydrogen bond donor; ahba, hydrogen bond acceptor.

Journal of Computational Chemistry

DOI 10.1002/jcc

Du et al. Vol. 29, No. 2 Journal of Computational Chemistry

218

the predictive power of QSAR technique promisingly. The PCA

IDLS method for solving the 3D simultaneous linear equation set
[eq. (7)] with the coefcients {ak} and {bj} are proved to be effective and successful with the calculation example. The converged
n
coefcients {ak } in Table 2 indicate that the ve potential elds
are not equally important. In Figure 3, the correlation coefcient
increases from R(0) 0.7895 to R(n) 0.9698, and the predicted
residue decreases from Q(0) 0.7534 to Q(n) 0.2916 with the
iterations. The R(0) and Q(0) are the results when the initial values
0
of {ak } are assigned to 1. Under such a condition the MF-3DQSAR is reduced to the standard 3D-QSAR. Accordingly, the correlation coefcient R(0) and the predictive residue Q(0) are actually
the results of the standard 3D-QSAR. On the other hand, the R(n)
and Q(n) are the improved results of MF-3D-QSAR. Similarly, in
Figure 4, the rst predicted activities are the predicted results of
traditional 3D-QSAR and the nally converged activities are the
predicted results of MF-3D-QSAR. All these remarkable improvea
Table 3. The pIC50 of 32 Molecules in Experiments, in Training
Calculations, and in Jackknife Tests.

Molecule
M01
M02
M03
M04
M05
M06
M07
M08
M09
M10
M11
M12
M13
M14
M15
M16
M17
M18
M19
M20
M21
M22
M23
M24
M25
M26
M27
M28
M29
M30
M31
M32
Rcalc 0.9627
Rpred 0.8194
a

pIC50_expta
6.650
7.220
9.520
7.920
7.660
9.000
9.000
6.210
6.280
8.000
9.000
9.000
8.520
6.220
6.740
6.680
6.700
6.570
5.200
8.050
9.000
5.430
7.220
5.700
6.890
6.740
6.520
6.700
7.050
6.280
5.660
5.700

pIC50_calcb

pIC50_predc

6.357
7.970
7.348
8.232
9.623
10.496
7.914
7.146
7.825
7.241
8.975
8.292
8.971
6.557
6.259
6.534
6.060
5.874
7.620
7.620
8.968
8.666
8.848
9.780
8.596
Unconverged
6.498
6.736
6.498
6.244
6.735
Unconverged
7.199
5.786
6.536
6.321
4.988
4.930
7.791
8.107
8.332
9.772
6.169
5.841
7.268
7.686
6.542
6.040
6.697
6.355
6.758
7.249
6.817
6.563
6.854
7.040
6.854
6.694
5.967
5.187
6.067
7.218
5.147
3.972
Qcalc 60.3226
Qpred 60.8578

From Maring et al.24

b
Calculated in training set of 32 molecules.
c
Predicted by jackknife test.

Figure 4. The convergence of predicted bioactivities of molecules

(M20, M21, M22) with the iterations in the jackknife test. In the
rst several iterations the predicted activities uctuate remarkably.
Then the predicted activities converge to the values close to the experimental data. The rst predicted activities are the results of traditional 3D-QSAR and the nally converged activities are the predicted results of MF-3D-QSAR. Promising improvements are from
the better theoretical model and the advanced mathematical technique of MF-3D-QSAR. [Color gure can be viewed in the online
issue, which is available at www.interscience.wiley.com.]

ments are from the better theoretical model and the advanced
mathematical technique of MF-3D-QSAR.
The bioactivities of drugs are affected by many factors,
including thermodynamic and nonthermodynamic factors. In
additional to the electrostatic and van der Waals interactions
(described by Coulomb law and Lennard-Jones equation) there
are many other factors. In MF-3D-QSAR, the potential elds are
not limited in the thermodynamic potentials; they may also
include other physical and chemical factors that affect either receptor recognition or activation of drug candidates. For further
improvement of MF-3D-QSAR, we need better linear free
energy equation, including more potential elds, and advanced
parametrization for the potential elds.

Acknowledgments
The authors wish to express their gratitude to the anonymous
reviewers whose constructive comments are very helpful for
strengthening the presentation of this paper. The work is supported by the Chinese National Basic Research Program (973)
under the project 2004CB719606, and by the Chinese National
Science Foundation (NSFC).

References
1. Hansch, C. Drug Design; Academic Press: New York, 1971; pp.
271275.
2. Hansch, C.; Leo, A. J. Substitute Constants for Correlation Analysis
in Chemistry and Biology; Wiley: New York, 1979.

Journal of Computational Chemistry

DOI 10.1002/jcc

Multiple Field 3D-QSAR

3. Hansch, C.; Streich, M.; Geiger, F.; Muir, R. M.; Maloney, P. P.;
Fujita, T. J Amer Soc 1963, 85, 2817.
4. Klopman, G.; Wang, S.; Jacobs, M. R.; Ellner, J. J. Antimicrob
Agents Chemother 1993, 37, 1807.
5. Klopman, G.; Wang, S.; Jacobs, M. R.; Bajaksouzian, S.; Edmonds,
K.; Ellner, J. J. Antimicrob Agents Chemother 1993, 37, 1799.
6. Klopman, G.; Fercu, D.; Renau, T. E.; Jacobs, M. R. Antimicrob
Agents Chemother 1996, 40, 2637.
7. Klopman, G.; Fercu, D.; Li, J. Y.; Rosenkranz, H. S.; Jacobs, M. R.
Res Microbiol 1996, 147, 86.
8. Cramer, R. D.; Patterson, D. E.; Bunce, J. D. J Am Chem Soc 1988,
110, 5959.
9. Cramer, R. D., III; Patterson, D. E.; Bunce, J. D. Prog Clin Biol Res
1989, 291, 161.
10. Du, Q.; Arteca, G. A. J Comput Aided Mol Des 1996, 10, 133.
11. Du, Q.; Arteca, G. A.; Mezey, P. G. J Comput Aided Mol Des
1997, 11, 503.
12. Du, Q.; Mezey, P. G. J Comput Aided Mol Des 1998, 12, 451.
13. Du, Q.; Liu, P.; Mezey, P. J Chem Inf Model 2005, 45, 347.
14. Du, Q.; Mezey, P.; Chou, K. C. J Comput Chem, 2005, 26, 461.
15. Du, Q.; Li, D.; He, W. Z.; Chou, K. C. J Comput Chem 2006, 27, 685.
16. Viswanadhan, V. N.; Ghose, A. K.; Hanna, N. B.; Matsumoto, S. S.;
Avery, T. L.; Revankar, G. R.; Robins, R. K. J Med Chem 1991, 34,
526.

17.
18.
19.
20.
21.
22.
23.
24.

25.
26.
27.
28.
29.
30.
31.
32.
33.

219

Klebe G.; Abraham U.; Mietzner T. J Med Chem 1994, 37, 4130.
Klebe, G.; Abraham, U. J Comput Aided Mol Des 1999, 13, 1.
Waller, C. L.; Marshall G. R. J Med Chem 1993, 36, 2390.
Jain, A. N.; Koile, K.; Chapman, D. J Med Chem 1994, 37, 2315.
Loew, G. H.; Villar, H. O.; Alkorta, I. Pharm Res 1993, 10, 475.
Kellogg, G. E.; Abraham, D. J. J Mol Graphics 1992, 10, 212.
Kim, K. H. Quant Struct-Act Relat 1993, 12, 232.
Maring, C. J.; Stoll, V. S.; Zhao, C.; Sun, M.; Krueger, A. C.; Stewart, K. D.; Madigan, D. L.; Kati, W. M.; Xu, Y.; Carrick, R. J.;
Montgomery, D. A.; Kempf-Grote, A.; Marsh, K. C.; Molla, A.;
Steffy, K. R.; Sham, H. L.; Laver, W. G.; Gu, Y. G.; Kempf, D. J.;
Kohlbrenner, W. E. J Med Chem 2005, 48, 3980.
Smith, B. J.; Colman, P. M.; Von Itzstein, M.; Danylec, B.; Varghese, J. N. Protein Sci 2001, 10, 689.
Wei, D. Q.; Du, Q. S.; Sun, H.; Chou, K. C. Biochem Biophys Res
Commun 2006, 344, 1048.
Wang, S. Q.; Du, Q. S.; Chou, K. C. Biochem Biophys Res Commun 2007, 354, 634.
Dunn, C. J.; Goa, K. L. Drugs 1999, 58, 761.
Lew, W.; Chen, X.; Kim, C. U. Curr Med Chem 2000, 7, 663.
Chou, K. C. Curr Med Chem 2004, 11, 2105.
Chou, K. C. Biochem Biophys Res Commun 2004, 316, 636.
Chou, K. C.; Zhang, C. T. Crit Rev Biochem Mol Biol 1995, 30, 275.
Chou, K. C.; Shen, H. B. J Prot Res 2007, 6, 1728.

Journal of Computational Chemistry

DOI 10.1002/jcc