November 16, 2007 17:48 WSPC/178-JTCC 00342

Journal of Theoretical and Computational Chemistry

Vol. 6, No. 4 (2007) 947–958

c World Scientific Publishing Company

QSAR AND ACTION MECHANISM OF TROXACITABINE

PRODRUGS WITH ANTITUMOR ACTIVITY

SI YAN LIAO, LIAN CAI XU, LI QIAN

and KANG CHENG ZHENG∗
School of Chemistry and Chemical Engineering
Zhongshan (Sun Yat-Sen) University
Guangzhou, 510275, P. R. China
∗ceszkc@mail.sysu.edu.cn

Received 12 July 2007

Accepted 22 August 2007

The quantitative structure–activity relationship (QSAR) of troxacitabine prodrugs with

antitumor activity has been studied by using the density functional theory (DFT), molec-
ular mechanics (MM2), and statistical methods. The established QSAR model shows not
only significant statistical quality, but also predictive ability, with the square of adjusted
correlation coefficient (R2A = 0.867) and the square of the cross-validation coefficient
(q 2 = 0.807). The antitumor activity is expressed as pIC 50 , which is defined as the neg-
ative value of the logarithm of necessary molar concentration of a compound to cause
50% growth inhibition against the human non-small-cell lung cancer cell line SW1573. It
appears to be mainly governed by two factors (or original variables), i.e. the calculated
hydrophobic coefficient (C log P ) of whole molecule and the net charges of the first atom
of substituent R (QFR ), although three descriptors, i.e. C log P , (C log P )2 , and QFR ,
were selected in our multiple linear regression model. The factor C log P shows parabolic
relation to pIC 50 and its suitable range is around 5.6, and the other factor QFR shows a
significant negative correlation with pIC50 . In this paper, a detailed discussion on these
two factors was carried out, and their close correlation with the action mechanism of
these prodrugs was reasonably revealed. Such results can offer some useful theoretical
references for understanding the action mechanism and directing the molecular design
of this kind of compound with antitumor activity.

Keywords: QSAR; action mechanism; troxacitabine; antitumor activity; DFT

calculation.

1. Introduction
Nucleoside analogues are among the first chemotherapeutic agents for the medical
treatment of cancer.1,2 These chemicals include several analogues of physiological
pyrimidine or purine nucleosides and nucleobases, which are thought to interact
with various intracellular targets involved in the metabolism of physiological nucle-
osides and DNA synthesis.3 Troxacitabine, (–)-2
-deoxy-3
-oxacytidine (Fig. 1), is a
novel unnatural L-nucleoside analogue with in vitro and in vivo antitumor activity
against human leukemias and a variety of solid tumors.4–6 It is phosphorylated to

947
November 16, 2007 17:48 WSPC/178-JTCC 00342

948 S. Y. Liao et al.

NH2

OH
O N

O
O

Fig. 1. Structural schematic diagram of troxacitabine.

active triphosphate form by intracellular kinases and produces cytotoxic activity

through incorporation of the triphosphate anabolite by DNA polymerases, leading
to DNA chain termination and inhibition of DNA replication.7,8 However, pharma-
cokinetic studies indicated that troxacitabine is slowly accumulated in cancer cells
by passive diffusion,9 and it is a hydrophilic agent and thus must be administered
intravenously in a frequent dosage schedule, which may result in greater toxicity
than a single dose schedule.10 So the development of new and more efficient ana-
logues is very significant.
Theoretically, improving the lipophilicity of prodrugs may benefit the intra-
cellular transport of compounds, making efficacy of drugs increase in cellular
uptake. Recently, Radi et al. have synthesized 20 troxacitabine prodrugs with
higher lipophilicity, and have evaluated their antitumor activities against two
non-small-cell lung cancer cell lines (A549 and SW1573).11 They found that
lipophilic troxacitabine prodrugs are indeed more active against solid tumors
in vitro than troxacitabine. However, the details of biochemical interactions for these
prodrugs remain unclarified. Fortunately, the availability of experimental cytotoxic
data of these prodrugs affords us an opportunity to apply quantitative structure–
activity relationship (QSAR) approach for obtaining some significant insights into
the action mechanism.
QSAR approach, which quantitatively correlates biological activities with the
molecule-structural descriptors, has been efficiently used for the study of biological
mechanisms of various reactive chemicals for many years.12 In order to establish
an excellent QSAR equation, it is very important to select high-quality descriptors,
because the success of a QSAR model is highly dependent on the choice of effective
descriptors. At present, a number of quantum-chemical descriptors like charges,
orbital energies, frontier orbital densities, and dipole moment, etc. can be easily
obtained by theoretical calculations. In particular, the Density Functional Theory
(DFT) calculations can clearly describe specific electronic properties and have been
successfully used in QSAR study.13,14 In addition, biological activities of molecules
are highly influenced not only by their inherent electronic properties, but also by
their transportabilities. The transport of compound through membranes can be
November 16, 2007 17:48 WSPC/178-JTCC 00342

QSAR and Action Mechanism of Troxacitabine Prodrugs with Antitumor Activity 949

experimentally measured by molecular hydrophobicity, which can be described by

octanol/water partition coefficient (C log P ).15 However, it is very important for
the QSAR study that such a descriptor as well as other descriptors with regard
to molecular properties can be obtained by the MM2 calculations. It makes QSAR
study more significant because all of the descriptors can be obtained via theoretical
calculations and do not depend on experiments.
In this work, a QSAR study of a series of troxacitabine prodrugs with antitumor
activity against human non-small-cell lung cancer cell line SW157311 was carried out
by using a combined method of the DFT, MM2, and statistics methods. A robust
and predictive QSAR equation only with two factors (or original variables), i.e. the
net charges of the first atom of substituent R (QFR ) and hydrophobicity C log P (in
our multiple linear regression model, three descriptors, i.e. C log P , (C log P )2 , and
QFR , were selected) was established. Based on this model, the action mechanism
was reasonably discussed, so this work can offer a useful insight into understanding
the action mechanism and directing the molecular design of this kind of compound
with antitumor activity.

2. Calculations
2.1. Computational methods
The molecular structures used in model building are given in Fig. 2. The molecular
mechanics MM2 method in Chem3D software16 was applied to search for lower
energy conformations for each molecule. Then the stable geometric structures of
these compounds were further fully optimized using the DFT calculations at the
level of Becke’s three-parameter hybrid functional (B3LYP),17 together with 6-31G
basis set. There was no imaginary frequency for all configurations at the energy
minima by the frequency calculations, showing that the optimized stable structures

HN R

OH
O N

O
O

Fig. 2. Structural schematic diagram of the troxacitabine prodrugs.

November 16, 2007 17:48 WSPC/178-JTCC 00342

950 S. Y. Liao et al.

are reasonable and reliable. Atomic charges were computed using the natural popu-
lation analysis (NPA).18 All of the quantum chemistry calculations were performed
by using the Gaussian 03 package of programs.19 The Hyperchem software20 was
also employed to calculate the following parameters of molecular properties from the
energy-minimized structures: surface area (S), volume (V ) of the whole molecule,
hydrophobic coefficient (log PR ), and molar refractivity (MRR ) of the substituent
(R), and so on. In addition, the calculated hydrophobic coefficient (C log P ) and
calculated molar refractivity (CMR) of the whole molecule were obtained by Chem-
Draw software.16

2.2. Statistical analysis

From the results of the DFT calculations, the quantum-chemical descriptors were
obtained for model building as follows: the highest occupied molecular orbital
energy (EHOMO ), the lowest unoccupied molecular orbital energy (ELUMO ), the
energy difference between the LUMO and the HOMO (∆EL-H ), the total dipole
moment (µ) of the molecule, the net charges of the first atom of substituent R
(QFR ), the total net charge of substituent R (ΣQR ), etc. Meanwhile, the descrip-
tors of molecular properties, such as surface area (S), volume (V ), the calcu-
lated hydrophobic coefficient (C log P ), the calculated molar refractivity coefficient
(CMR), were also selected. More than 20 descriptors were adopted as candidates
for the correlation analysis.
The antitumor activities of 20 studied compounds are expressed in terms of
IC50 , which is necessary molar concentration of a compound causing 50% cell
growth inhibition against the human non-small-cell lung cancer cell line SW1573;
the corresponding values are presented in Table 1.11 All original IC50 values were
converted to negative logarithm of IC50 (pIC50 ) in the QSAR study.
To select out the predominant descriptors affecting the cytotoxic activity, the
correlation analysis was performed by the statistical software SPSS.21 To eliminate
the inter-correlative parameters and minimize the information overlap in the model,
the descriptors with higher intercorrelation (|r| > 0.5) were not considered.22 The
descriptors with higher correlation to the pIC50 and lower intercorrelation were
selected to carry out the stepwise multiple linear regression analysis to establish
the optimal QSAR equation.
In the following step, the predictive ability of the equation was evaluated by
a well-known method of “leave-one-out” (LOO) cross-validation.23 The square of
cross-validated coefficient q 2 , which is used as a criterion of both robustness and
predictive ability of the model, should be larger than 0.5 for a reliable model.24
It is commonly accepted that the internal validation of QSAR model built from
training set is sufficient to confirm its predictive power. However, a high q 2 in
training set only shows a good internal validation, but it does not automatically
infer its high predictive ability for an external test set.25,26 To obtain QSAR with
more reliable predictive ability, external validation is also crucial. Thus, the data
November 16, 2007 17:48 WSPC/178-JTCC 00342

QSAR and Action Mechanism of Troxacitabine Prodrugs with Antitumor Activity 951

Table 1. Structures and experimental antitumor activities

of the troxacitabine prodrugs (against human non-small-
cell lung cancer cell line SW1573).

Compound Substituent R IC50 (µM)

Troxacitabine — 2.3
6a CH3 10.70
6b (CH2 )2 CH3 8.07
6c (CH2 )3 CH3 9.13
6d (CH2 )5 CH3 2.40
6ea (CH2 )7 CH3 0.276
6f (CH2 )8 CH3 0.040
6g (CH2 )10 CH3 0.03
6h (CH2 )14 CH3 0.020
6ia (CH2 )16 CH3 0.125
6j (CH2 )18 CH3 0.150
6k (CH2 )20 CH3 0.900

6l 22.00

6ma 8.80

6n 14.00

6o 13.67

6p F 15.33

6qa Cl 2.77

6r Br 8.87

6s OMe 10.03

Cl
6t 14.33

Cl

a Compounds in the test set.

November 16, 2007 17:48 WSPC/178-JTCC 00342

952 S. Y. Liao et al.

of 20 compounds were divided into a training set comprised of randomly selected

16 compounds and a test set comprised of the rest of the four compounds. The
optimal QSAR model was built from training set, and it was determined with the
test set to confirm its predictive ability.

3. Results and Discussion

3.1. QSAR equations
Considering the balance of the QSAR quality and the number of employed descrip-
tors, an optimal QSAR equation was obtained for 16 compounds in training set
through stepwise multiple linear regression analysis as follows:
pIC50 = 4.347 − 0.071(C log P )2 + 0.718C log P − 2.177QFR
n = 16 R2 = 0.893 RA
2
= 0.867 r = 0.345 Sreg = 0.535 Spre = 0.463 (1)
2
q = 0.807 F = 33.49 p < 0.0001
where n is the number of compounds in training set; R2 is the square of correlation
2
coefficient of regression; RA is the square of adjusted correlation coefficient; r is the
root-mean-square deviation; Sreg is the standard deviation of regression and Spre is
that of the prediction; F is the Fisher’s F -value using the F statistics; p is the p-
value using the F statistics; and q 2 is the square of LOO cross-validation coefficient.
A good QSAR model has characters of large F , small r and S, low p-value, R2 and
q 2 values close to 1. In general, the QSAR model is significant at p < 0.001. So
above established QSAR model Eq. (1) shows a significant statistical quality, not
2
only in a reliability (RA = 0.867) but also in a predictability (q 2 = 0.807).
It is very interesting that, Eq. (1) as a multiple linear regression model owns
three descriptors ((C log P )2 , C log P , and QFR ) but only two factors or original
variables (C log P and QFR ). Such an equation not only reflects the multiple linear
relationship between the activity and the three descriptors, but also reflects the
unlinear (parabolic) relationship between the activity and C log P . The former is
advantageous to establishing QSAR equation according to the multiple linear model
and the latter is advantageous to discussing the mechanism as shown in Sec. 3.2. In
the multiple linear regression analysis, C log P and (C log P )2 as candidate descrip-
tors were simultaneously selected in Eq. (1) through eliminating the descriptors
with higher inter-correlation (here |r| > 0.5). If the criteria of inter-correlation is
adjusted to a higher level, more descriptors may be selected to this model. For
example, applying the criteria |r| > 0.7, Eq. (2) can be obtained as follows:
pIC50 = 18.99 − 0.072(C log P )2 + 0.727C log P − 2.730QFR + 59.89EHOMO
n = 16 R2 = 0.905 RA
2
= 0.871 r = 0.325 Sreg = 0.549 Spre = 0.456 (2)
2
q = 0.814 F = 26.24 p < 0.0001.
Comparing Eq. (2) with Eq. (1), we find that Eq. (2) shows a little improvement
only in larger R2 , RA
2
, and q 2 values as well as smaller r and S values, but an
November 16, 2007 17:48 WSPC/178-JTCC 00342

QSAR and Action Mechanism of Troxacitabine Prodrugs with Antitumor Activity 953

additional descriptor EHOMO is selected and a smaller F value is caused. Such

results show that Eq. (2) does not have substantial improvement. On the contrary,
it is not good in the statistical quality, because according to the Topliss and Costello
rule,27 in order to minimize the risk of chance correlations, the ratio of number of
compounds to that of descriptors should reach at least 5:1, whereas Eq. (2) does
not reach this standard (only 4:1). So the criteria of inter-correlation with |r| >0.5
is feasible, Eq. (1) is optimal one in this work and we will only discuss it below.
The selected descriptors in Eqs. (1) and (2) as well as the deviations of regression
and prediction are listed in Table 2.
The correlation matrix between pIC50 and selected descriptors in Eqs. (1) and
(2) are shown in Table 3, it shows that the selected descriptors in Eq. (1), i.e. QFR ,
C log P , and (C log P )2 , are independent with small intercorrelation coefficients
(|r| < 0.4) except for a high intercorrelation coefficient between (C log P )2 and
C log P (R = 0.939), which is acceptable because the values of (C log P )2 are cal-
culated from C log P . These results show that Eq. (1) is of an excellent robustness.
A well-accepted QSAR model should be able to accurately predict activities of
new compounds. Therefore, the QSAR model Eq. (1) was further applied to predict
the antitumor activities of four compounds in the test set. From Table 4, we can
find that the predicted deviations lie in a range of −0.45 to 0.26, and the regression
deviations are within the range of −0.64 to 0.74. Hence, the results obtained from
the test set further demonstrate that this QSAR model has an excellent predictive
ability. The plot of the predicted pIC50 values versus experimental ones based on
Eq. (1) is shown in Fig. 3. Obviously, the predicted pIC50 values are in a good
agreement with experimental ones.

3.2. Discussion on the QSAR and related action mechanism

It is very interesting to find that this optimal QSAR model Eq. (1) only includes two
factors (or original variables) expressing different properties, that is to say, trans-
portability property (C log P ) and inherent electronic property (QFR ), although
it has three descriptors in the multiple linear regression analysis. It is these two
factors that are most responsible for the antitumor activity (pIC50 ) of the studied
compounds.
The evaluation of the correlation coefficient of C log P versus pIC50 (R = 0.723)
shows that C log P has the highest contribution to the pIC50 ; it implies that in vitro
cytotoxicities of these compounds mainly depend on their transportability property.
In Fig. 4, antitumor activity (pIC50 ) shows a parabolic relation to the parameter
C log P , in which the pIC50 increases with the increase in C log P value up to a
critical point (optimal C log P value being ∼5.6), and then the pIC50 decreases.
That is to say, the compounds with very low or very high hydrophobicity are
disadvantageous to improving their antitumor activities; the compounds with a
suitable range of hydrophobicity or lipophilicity (C log P of ∼5.6) should have a
potent antitumor activity. In Table 2, we can see that compound 6h with optimal
 954

Table 2. Calculated results for 16 compounds in the training set using Eqs. (1) and (2).

Reg. Dev.a Pred. Dev.b Reg. Dev.a Pred.

S. Y. Liao et al.

for Eq. (1) for Eq. (1) for Eq. (2) Dev.b for
Comp. (C log P )2 C log P QFR (a.u.) EHOMO (a.u.) pIC50 (expt.) Eq. (2)
6a 2.587 −1.609 −0.7974 −0.2512 4.971 −0.23 −0.53 −0.20 −0.49
6b 0.3031 −0.5506 −0.5671 −0.2500 5.093 0.07 0.08 0.05 0.06
6c 0.0005 −0.0216 −0.5644 −0.2499 5.04 0.52 0.60 0.51 0.59
6d 1.074 1.036 −0.5634 −0.2497 5.62 0.62 0.74 0.63 0.75
6f 6.882 2.624 −0.5634 −0.2496 7.398 −0.43 −0.58 −0.41 −0.55
November 16, 2007 17:48 WSPC/178-JTCC

6g 13.55 3.682 −0.5635 −0.2495 7.523 −0.27 −0.38 −0.24 −0.34

6h 33.61 5.798 −0.5634 −0.2497 7.699 −0.35 −0.47 −0.33 −0.45
6j 62.62 7.914 −0.5634 −0.2496 6.824 −0.01 −0.03 0.00 −0.02
6k 80.49 8.972 −0.5634 −0.2496 6.046 0.25 0.68 0.26 0.66
6l 1.574 4.658
00342

−1.255 −0.3815 −0.2492 −0.49 −0.64 −0.58 −0.79

6n 0.1785 0.4224 −0.3485 −0.2492 4.854 0.54 0.59 0.46 0.52
6o 0.0143 −0.1196 −0.1606 −0.2438 4.864 −0.25 −0.31 −0.12 −0.17
6p 0.0023 0.0481 −0.1707 −0.2476 4.814 −0.06 −0.07 −0.15 −0.19
6r 0.5900 0.7681 −0.1603 −0.2478 5.052 0.15 0.19 0.05 0.07
6s 0.0007 −0.0256 −0.1873 −0.2403 4.999 −0.26 −0.31 0.09 0.39
6t 0.2599 0.5098 −0.1663 −0.2491 4.844 0.21 0.26 0.03 0.05

a The regression (reg.) deviations (dev.) for all compounds, reg. dev. = pIC50 (reg.) −pIC50 (expt.).
b The prediction (pred.) deviation (dev.) for compound i(i = 1, . . . , n), pred. dev. = pIC 50 (pred.)i − pIC50 (expt.)i , here pIC50 (pred.)i is
calculated by the reg. equation obtained after leaving compound i out.
November 16, 2007 17:48 WSPC/178-JTCC 00342

QSAR and Action Mechanism of Troxacitabine Prodrugs with Antitumor Activity 955

Table 3. Correlation matrix for pIC50 and selected parameters in

Eqs. (1) and (2).

pIC50 (C log P )2 C log P QFR EHOMO

pIC50 1.000
(C log P )2 0.516 1.000
C log P 0.723 0.939 1.000
QFR −0.486 −0.356 −0.309 1.000
EHOMO −0.290 −0.229 −0.217 0.684 1.000

Table 4. Computational results for four compounds in the test set using Eq. (1).

Compound (C log P )2 C log P QFR1 (a.u.) pIC50 (expt.) pIC50 (calc.)a Devb
6e 4.387 2.094 −0.5636 6.559 6.766 0.21
6i 47.00 6.856 −0.5634 6.903 7.159 0.26
6m 0.0186 −0.1366 −0.3583 5.056 5.028 −0.03
6q 0.3821 0.6181 −0.1601 5.558 5.112 −0.45

a Predictive activities were calculated using Eq. (1).

b Dev. = pIC50 (calc.) −pIC50 (expt.).

7.5

7.0

6.5
Predicted pIC50

6.0

5.5

5.0

4.5

4.0

4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5

Experimental pIC50

Fig. 3. Plot of predicted activities versus experimental ones based on Eq. (1), in which 16 com-
pounds in the training set are expressed by dots and four compounds in the test set are expressed
by triangles.

C log P value (5.798) possesses the highest activity (pIC50 = 7.699) among this
data set, whereas compound 6a with smaller C log P value (−1.609) has lower activ-
ity (pIC50 = 4.971); compound 6k with larger C log P value (8.972) also has lower
activity (pIC50 = 6.046). Such a result suggests that a very low hydrophobicity may
November 16, 2007 17:48 WSPC/178-JTCC 00342

956 S. Y. Liao et al.

8.0

7.5

7.0
Experimental pIC50

6.5

6.0

5.5

5.0

4.5

-2 0 2 4 6 8 10
ClogP

Fig. 4. Plot of experimental activities (pIC50 ) versus C log P values.

restrict the transport of pharmaceutical molecule into active body tissues through
membranes, whereas a very high hydrophobicity may be disadvantageous to their
hydrolysis processes, which are necessary procedures resulting in a ultima specimen
with anticancer activity for these prodrugs.11 In addition, the diffusion of dots in
Fig. 4 implies that there are also other fators influencing the cytotoxicity.
On the other hand, the electronic parameter QFR exhibits negative correla-
tion coefficient with pIC50 (R = −0.486), so the more negative the QFR value,
the higher the activity of the compound is. Such a regularity can be easily found
through comparisons between compounds 6b and 6c, 6n and 6o, and other ones.
The more negative charges at the first atom of substituent R mean the stronger
electron-withdrawing ability of this atom and then make the positive charge of C
atom of adjacent carbonyl increase, so it must be advantageous to the nucleophilic
attack of water molecules for this C atom in active body tissues, and thus advan-
tageous to its hydrolysis processes resulting in a ultima specimen with anticancer
activity. Such a result is in good agreement with the hypothesis mechanism pro-
posed by Radi et al. who suggested that the amidase-catalyzed hydrolysis of the
prodrugs might play an important role, allowing the release of troxacitabine into
the intracellular compartment for triphosphate formation.11 In this QSAR model,
the selected descriptor QFR is closely correlative with the anticancer activity of
these compounds, and thus further supports Radi’s hydrolysis action mechanism of
the prodrugs, at least, in the preliminary procedure.
November 16, 2007 17:48 WSPC/178-JTCC 00342

QSAR and Action Mechanism of Troxacitabine Prodrugs with Antitumor Activity 957

4. Conclusions
The QSAR of this series of troxacitabine prodrugs with antitumor activity against
non-small-cell lung cancer cell line SW1573 has been studied by using the DFT,
MM2, and statistical methods. It is very interesting to find that the cytotoxicity
of this kind of compound mainly depends on two factors: the calculated hydropho-
bic coefficient (C log P ) and the net charge variable (QFR ) of the first atom of
substituent R. The established QSAR shows not only acceptable regressive perfor-
2
mance (RA = 0.867) but also considerable predictive ability (q 2 = 0.807). Moreover,
the predicted pIC50 values for the four compounds in test set are in a considerable
agreement with corresponding experimental ones. Such a result further confirms
this established QSAR having excellent predictive ability. Based on this QSAR
model, some useful insights are obtained as follows: (1) considering a suitable range
of lipophilicity (C log P of ca. 5.6) is very important to obtain the higher antitumor
activities for this series of prodrugs. (2) Selected another descriptor, i.e. the net
charges of the first atom of substituent R (QFR ) further supports the hydrolysis
action mechanism of the prodrugs, at least, in the preliminary procedure. These
theoretical results can be used as references for understanding the action mecha-
nism and directing the molecular design of this kind of compound with antitumor
activity.

Acknowledgments
The financial supports of the National Natural Science Foundation of China and
the Natural Science Foundation of Guangdong Province of China are gratefully
acknowledged.

References
1. Jordheim L, Galmarini CM, Dumontet C, Lancet Oncol 3:415, 2002.
2. Jordheim L, Galmarini CM, Dumontet C, Curr Drug Targets 4:443, 2003.
3. Helguera AM, Rodrı́guez-Borges JE, Garcı́a-Mera X, Fernández F, Cordeiro
MNatáliaDS, J Med Chem 50:1537, 2007.
4. Gourdeau H, Bibeau L, Ouellet F, Custeau D, Bernier L, Bowlin T, Cancer Chemother
Pharmacol 47:236, 2001.
5. Rabbani SA, Harakidas P, Bowlin T, Attardo G, Cancer Res 58:3461, 1998.
6. Weitman S, Marty J, Jolivet J, Locas C, Von Hoff DD, Clin Cancer Res 6:1574, 2000.
7. Kukhanova M, Liu SH, Mozzherin D, Lin TS, Chu CK, Cheng YC, J Biol Chem
270:23055, 1995.
8. Gunzburg WH, Salmons B, Trends Mol Med 7:30, 2001.
9. Gourdeau H, Clarke ML, Ouellet F, Mowles D, Selner M, Richard A, Lee N, Mackey
JR, Young JD, Jolivet J, Lafreniere RG, Cass CE, Cancer Res 61:7217, 2001.
10. De Bono JS, Stephenson J, Baker SD, Hidalgo M, Patnaik A, Hammond LA, Weiss
G, Goetz A, Siu L, Simmons C, Jolivet J, Rowinsky EK, J Clin Oncol 20:96, 2002.
11. Radi M, Adema AD, Daft JR, Cho JH, Hoebe EK, Alexander LEMM, Peters GJ,
Chu CK, J Med Chem 50:2249, 2007.
12. Isayev O, Rasulev B, Gorb L, Leszczynski J, Mol Divers 10:233, 2006.
November 16, 2007 17:48 WSPC/178-JTCC 00342

958 S. Y. Liao et al.

13. Chen JC, Qian L, Wu WJ, Chen LM, Zheng KC, J Mol Struct Theochem 756:167,
2005.
14. Wu WJ, Chen JC, Qian L, Zheng KC, J Theor Comput Chem 2:223, 2007.
15. Lameira J, Alves CN, Moliner V, Silla E, Eur J Med Chem 41:616, 2006.
16. CambridgeSoft Corp., 100 Cambrige Park, MA, 02140-2317, USA.
17. Becke AD, J Chem Phys 98:1372, 1993.
18. Carpenter JE, Weinhold F, J Mol Struct Theochem 169:41, 1988.
19. Frisch MJ, Trucks GW, Schlegel HB, Scuseria GE, Robb MA, Cheeseman JR, Mont-
gomery JA, Vreven TJr, Kudin KN, Burant JC, Millam JM, Iyengar SS, Tomasi J,
Barone V, Mennucci B, Cossi M, Scalmani G, Rega N, Petersson GA, Nakatsuji H,
Hada M, Ehara M, Toyota K, Fukuda R, Hasegawa J, Ishida M, Nakajima T, Honda
Y, Kitao O, Nakai H, Klene M, Li X, Knox JE, Hratchian HP, Cross JB, Bakken V,
Adamo C, Jaramillo J, Gomperts R, Stratmann RE, Yazyev O, Austin AJ, Cammi R,
Pomelli C, Ochterski JW, Ayala PY, Morokuma K, Voth GA, Salvador P, Dannen-
berg JJ, Zakrzewski VG, Dapprich S, Daniels AD, Strain MC, Farkas O, Malick DK,
Rabuck AD, Raghavachari K, Foresman JB, Ortiz JV, Cui Q, Baboul AG, Clifford S,
Cioslowski J, Stefanov BB, Liu G, Liashenko A, Piskorz P, Komaromi I, Martin RL,
Fox DJ, Keith T, Al-Laham MA, Peng CY, Nanayakkara A, Challacombe M, Gill
PMW, Johnson B, Chen W, Wong MW, Gonzalez C, Pople J, Gaussian 03, Revision
D.1, A. Gaussian, Inc., Wallingford, CT, 2005.
20. HyperChem, release 6.03; Hypercube Inc., http://www.hyper.com/, 2001.
21. SPSS, vs 9.0; SPSS Inc., Chicago, IL.
22. De K, Sengupta C, Roy K, Bioorg Med Chem 12:3323, 2004.
23. Tetko IV, Tanchuk VY, Villa AE, J Chem Inf Comput Sci 41:1407, 2001.
24. Golbraikh A, Tropsha A, J Mol Graph Model 20:269, 2002.
25. Zhang S, Wei L, Bastow K, Zheng W, Brossi A, Lee KH, Tropsha A, J Comput Aided
Mol Des 21:97, 2007.
26. Roy DR, Sarkar U, Chattaraj PK, Mitra A, Padmanabhan J, Parthasarathi R, Sub-
ramanian V, Van Damme S, Bultinck P, Mol Divers 10:119, 2006.
27. Topliss JG, Costello RJ, J Med Chem 10:1066, 1972.