ANN-QSAR model for selection of anticancer leads from structurally heterogeneous series of compounds

Article abstract of DOI:10.1016/j.ejmech.2006.11.016

Developing a model for predicting anticancer activity of any classes of organic compounds based on molecular structure is very important goal for medicinal chemist. Different molecular descriptors can be used to solve this problem. Stochastic molecular descriptors so-called the MARCH-INSIDE approach, shown to be very successful in drug design. Nevertheless, the structural diversity of compounds is so vast that we may need non-linear models such as artificial neural networks (ANN) instead of linear ones. SmartMLP-ANN analysis used to model the anticancer activity of organic compounds has shown high average accuracy of 93.79% (train performance) and predictability of 90.88% (validation performance) for the 8:3-MLP topology with different training and predicting series. This ANN model favourably compares with respect to a previous linear discriminant analysis (LDA) model [H. Gonzalez-Diaz et al., J. Mol. Model 9 (2003) 395] that showed only 80.49% of accuracy and 79.34% of predictability. The present SmartMLP approach employed shorter training times of only 10 h while previous models give accuracies of 70-89% only after 25-46 h of training. In order to illustrate the practical use of the model in bioorganic medicinal chemistry, we report the in silico prediction, and in vitro evaluation of six new synthetic tegafur analogues having IC₅₀ values in a broad range between 37.1 and 138 μg mL^-1 for leukemia (L1210/0) and human T-lymphocyte (Molt4/C8, CEM/0) cells. Theoretical predictions coincide very well with experimental results.

Full text of DOI:10.1016/j.ejmech.2006.11.016

European Journal of Medicinal Chemistry 42 (2007) 580e585
http://www.elsevier.com/locate/ejmech
Original article
ANN-QSAR model for selection of anticancer leads from
structurally heterogeneous series of compounds
a,*
b,c
d
e
Humberto Gonz a´ lez-D ´ı az , Isis Bonet , Carmen Teran , Erik De Clercq ,
´
b
b
a
a
Rafael Bello , Maria M. Garc ´ı a , Lourdes Santana , Eugenio Uriarte
a
Department of Organic Chemistry, University of Santiago de Compostela, 15782 Santiago de Compostela, Spain
b
CEI, UCLV, Santa Clara 54830, Cuba
c
Department of Plant Systems Biology, Flanders Interuniversity Institute for Biotechnology (VIB), Ghent University, Belgium
d
Department of Organic Chemistry, University of Vigo, 36200 Vigo, Spain
e
Katholieke Universiteit Leuven, Rega Institute, Minderbroedersstraat 10, B-3000 Leuven, Belgium
Received 16 September 2006; received in revised form 29 November 2006; accepted 30 November 2006
Available online 15 December 2006
Abstract
Developing a model for predicting anticancer activity of any classes of organic compounds based on molecular structure is very important
goal for medicinal chemist. Different molecular descriptors can be used to solve this problem. Stochastic molecular descriptors so-called the
MARCH-INSIDE approach, shown to be very successful in drug design. Nevertheless, the structural diversity of compounds is so vast that
we may need non-linear models such as artificial neural networks (ANN) instead of linear ones. SmartMLP-ANN analysis used to model
the anticancer activity of organic compounds has shown high average accuracy of 93.79% (train performance) and predictability of 90.88%
(
validation performance) for the 8:3-MLP topology with different training and predicting series. This ANN model favourably compares with
respect to a previous linear discriminant analysis (LDA) model [H. Gonz a´ lez-D ´ı az et al., J. Mol. Model 9 (2003) 395] that showed only
0.49% of accuracy and 79.34% of predictability. The present SmartMLP approach employed shorter training times of only 10 h while previous
8
models give accuracies of 70e89% only after 25e46 h of training. In order to illustrate the practical use of the model in bioorganic medicinal
chemistry, we report the in silico prediction, and in vitro evaluation of six new synthetic tegafur analogues having IC values in a broad range
5
0
ꢀ
1
between 37.1 and 138 mg mL for leukemia (L1210/0) and human T-lymphocyte (Molt4/C8, CEM/0) cells. Theoretical predictions coincide
very well with experimental results.
Ó 2006 Elsevier Masson SAS. All rights reserved.
Keywords: QSAR; Anticancer activity; ANN; Tegafur analogues; Linear discriminant analysis
1
. Introduction
invariant c, and Balaban index J may be expressed as vec-
torematrixevector forms. More recently, other interesting
topologic indices such as the so-called Marrero-Ponce qua-
Quantitative structureeactivity relationships (QSARs) have
emerged as a rational alternative in order to find new active
molecules including anticancer compounds [1,2]. Many topo-
logical molecular descriptors can be used to describe organic
molecular structure with QSAR aims. Almost of them, for
instance the Wiener index W, Harary number H, Randic
dratic indices q (X ) have been introduced [3e7]. In particular,
k
spectral moments of different matrices have been of special
relevance for QSAR and other studies such as Gutman and
Rosenfield studies on polymers’ graphs, mean hydrophobicity
moment, and the I3 index for proteins [8,9]. The success of the
method of moments on the field of polymers has also been
confirmed after G o´ nzalez M.P., et al. work on the analysis
of the mutagenic power of dental polymers [10], which pre-
ceded the very related and recent work of Morales, et al.
[11]. In general, the method of spectral moments has been
*
Corresponding author.
E-mail addresses: gonzalezdiazh@yahoo.es, qohumbe@usc.es (H. Gonz a´ lez-
D ´ı az).
0223-5234/$ - see front matter Ó 2006 Elsevier Masson SAS. All rights reserved.
doi:10.1016/j.ejmech.2006.11.016

H. Gonz a´ lez-D ´ı az et al. / European Journal of Medicinal Chemistry 42 (2007) 580e585
581
largely used in many other different contexts of solid, theo-
retic, and bioorganic chemistry [12e14]. G o´ nzalez M.P. has
additionally reported interesting applications of the method
of moment in the molecular design of herbicides [15]. Other
interesting applications of the method of moments in pharma-
ceutical sciences and medicinal chemistry were reported by
Cabrera-P e´ rez et al. [16], and Molina et al., design of antibi-
otics [17]. Last but not least, several applications on medicinal
chemistry of the method of moments developed by Estrada
and Pe n˜ a have appeared, including the design of sedative/hyp-
notic compounds and on the design of anticonvulsant drugs
techniques or a learning coefficient. In order to speed up and
optimize the ANN training process, SmartMLP version 1.5 en-
hances heuristics implemented in older versions. The learning
coefficient was made a function on the size of the training set
and the size of the samples in each class. Patterns selection
heuristic over training set was improved gathering together
some techniques implemented in older versions. The training
algorithm was modified to use both uniform and RUL tech-
niques in a way that it starts applying RUL until an error
threshold was reached (e < 0.1) and then it applies uniform
selection [32,33].
[
18].
Particularly, our group has worked on a Markov model that
In order to optimize the performance of the method we
trained the ANN using different topologies with one or two
hidden layers and a variable number of neurons. SmartMLP
1.5 exhibited the best differentiation of anticancer from non-
active compounds with 94.04% of accuracy. This topology
presented two hidden layers having eight and three neurons,
respectively, see Table 2. Changing network topology is
a way to improve the QSAR results [34].
SR
uses stochastic spectral moments pk to encode molecular
structure with applications in nucleic acids, proteins and
medicinal chemistry research. In any case, it has been demon-
strated using vectorematrixevector formalism that all these
indices, including spectral moments too, are in fact very sim-
ilar [19e25].
On the other hand, the search of anticancer compounds has
always been on the desktop of molecular modelling and drug
design specialists. In spite of this intensive search, the discov-
ery of selective antitumor compounds has remained a largely
elusive goal of cancer research. Subsequently, new approaches
are needed in order to make an efficient search for candidates
to be assayed as anticancer drugs [26e31]. Further, in the
present work we have compared the LDA analysis previously
reported by our group [26] with new one based on an ANN
approach [32e34] using the database of 961 chemical com-
pounds. In brief, first we have calculated 11 stochastic spectral
Finally, we have tested the predictability and robustness of
the present ANN model by a re-substitution approach, which
consists of different data partition in interchangeable training
and predicting series [35]. SmartMLP 1.5 showed for all parti-
tions an average training accuracy of above 93.79% and similar
mean predictability of 90.88% in predicting series. Conversely,
the LDA model previously reported suffers a lack of accuracy
and predictability from 90.46% and 86.07% (partition 3) to av-
erages of 80.49% and 79.34%, respectively, after re-substitu-
tion approach. These results have shown without doubt, the
higher robustness of the present ANN models with respect to
SR
SR
moments pk for each compound, followed by training and
the LDA one [35], see Table 3. The names, pk values,
validation of new ANN using this data and finally testing
the newly synthesized molecules selected using the derived
model in order to prove the usefulness of model.
observed, and predicted classification of all the compounds
used to train and test the model appears in supplementary ma-
terial file, where the compounds misclassified are highlighted
in red.
2
. Results and discussion
2.2. An example on the use of the model
2
.1. Model train and validation
The ANN model developed was used to predict the biolog-
ical activity of some cyclopentylpirimidine derivative ana-
logues of tegafur, a largely known anticancer drug (Fig. 1).
Table 1 depicts experimental results obtained from training
a network by several types of MLP, with their time of training
and the learning percentage. The results show clearly that
SmartMLP version 1.5 increases learning rate while decreases
required training time even if we use combined selection
Table 2
SmartMLP classification percent for different topologies of one and two
hidden layers
a
Topologies
SmartMLP 1.5
accuracy (%)
Table 1
Variants of MLP, time of training and learning percent
One hidden layer
Two hidden layers
2
5
70.98
85.30
Backpropagation type
Training time (h)
Learning (%)
Traditional
46:50:00
30:00:00
25:00:00
25:00:00
70
82
89
87
6
8
87.58
88.60
SmartMLP version 1.0
Using learning coefficient
Combining selection techniques
First layer
Second layer
5
6
2
3
90.50
91.40
(
RUL and Uniform)
Combining both previous
SmartMLP version 1.5)
10:00:00
94.04
(
8
8
2
3
93.12
94.04
SmartMLP helped selecting an adequate network topology from several
candidates.
a
Number of neurons in the hidden layers.

582
H. Gonz a´ lez-D ´ı az et al. / European Journal of Medicinal Chemistry 42 (2007) 580e585
Table 3
Comparison of ANN and LDA classification results for several partitions of
the dataset
O
N
OTMS
N
R
Br
NH
R
a
Sets
SmartMLP 1.5
learning (%)
LDA
accuracy (%)
+
O
N
OTMS
Partition 1
Partition 2
Partition 3
Partition 4
Mean
Training set
Test set
93.48
91.67
77.5
78.3
1
2
3
R = H
R = CH3
R = F
Training set
Test set
93.62
90.83
77
76.3
Training set
Test set
94.01
89.75
90.46
86.07
Fig. 2. Reagents and conditions: (a) i: CH
2
Cl
2
, reflux; ii: MeOH/H
2
O (6:1).
Training set
Test set
94.04
91.25
77
76.7
compounds from heterogeneous series. This new strategy
showed higher accuracy and robustness than LDA-based strat-
egies. MARCH-INSIDE and SmartMLP-ANN approach was
able to predict anticancer activity on leukemia and lymphocyte
lines for pyrimidine derivatives, experimentally confirmed. All
these conclusions coincide with multiple applications of Mar-
kov process in the literature [38e42].
Training set
Test set
93.79
90.88
80.49
79.34
We predicted the biological activity only for simple analogues
from which CH , F, Cl, Br, and I substituted at the 5-carbon of
3
the uracil skeleton was selected (Figs. 2 and 3). This selection
was based on the simplicity and synthetic accessibility. All
these analogues were synthesized in enough quantity for bio-
logical assay. Compounds 1e3 were prepared by condensation
of the trimethylsilylated base (uracil, thymine, and fluoroura-
cil) with bromocyclopentane in 11%, 10%, and 13% yield, re-
spectively (Fig. 2) [36,37]. Compounds 4e6 were prepared
from 1 using N-chlorosuccinimide or N-bromosuccinimide in
4. Experimental section
4.1. The stochastic spectral moments
A precise definition of the descriptors generated by this
methodology can be found in several reports of its application
in the study of several biological properties. Briefly, the
method uses stochastic or Markov matrix P as a source of
molecular descriptor. The Markov matrix is built up as
1
acetic acid, or I in nitric acid and dioxane, with 68%, 93%,
2
and 94% yield, respectively (Fig. 3) [36].
The predictions coincide with the biological test result
where all the compounds showed detectable biological activity
in the three studied cellular lines namely L1210/0, Molt4/C8,
and CEM/0 (Table 4). The most interesting activity on the leu-
kemia line L1210/0 was presented by the F-tegafur analogue
a squared matrix n ꢁ n (n number of atoms in the molecule)
1
whose elements ( pij) are calculated as the ratio between the
Pauling’s electronegativity (c
of the c values for all the atoms covalently linked to the ith
) of the jth atom and the sum
j
k
atom, including itself [40].
(
compound 3). However, results in the human lymphocyte
u
cj,e
lines were more discrete. Interestingly, the Br-tegafur ana-
logue (compound 5) resulted more selective for the human
lymphocyte lines than the leukemia line. This result confirms
the potentialities of the MARCH-INSIDE approach to model
biological data and guide drug discovery in bioorganic medic-
inal chemistry [38,39].
1
p ¼ P
ð1Þ
ij
dþ1
u
ck,e
k¼1
Local 3D characteristics of each atom are codified through-
out the dummy variable u . This variable (u ) takes the value
j
j
u ¼ 1 if the atom a is R, E or axial and takes the values
j
j
u ¼ 0, ꢀ1 whether the atom has no specific 3D properties
j
or is S, Z or equatorial. The symbols R, S refer to the chirality
of the atom. Alternatively, ZeE regards to the 3D
3
. Conclusions
We can conclude that the combination of MARCH-INSIDE
with SmartMLP-ANN correctly classifies several anticancer
O
N
O
N
R
I
NH
NH
a
b
O
1
O
O
F
NH
N
O
4
R = Cl
R = Br
6
5
O
Fig. 3. Reagents and conditions: (a) NXS (X ¼ Cl, Br), AcOH, reflux; (b) I
2
,
ꢂ
Fig. 1. Tegafur.
HNO
3
, dioxane, 100 C.

H. Gonz a´ lez-D ´ı az et al. / European Journal of Medicinal Chemistry 42 (2007) 580e585
583
Table 4
Predicted, observed class and IC50 values for tegafur derivatives
1 Uniform: In this selection type, each pattern is selected
randomly but the probability of being selected is the
same one for each pattern.
2 Sequentially: The examples here are selected in the same
order that appear in the training set.
3 Repeat until learning (RUL): This is a pedagogic selection
type strategy. Each example is presented to the network
depending on the error the network commits classifying
it. So that each example is selected randomly and repeated
until its error is below the error average made by the net-
work increased by a certain factor b.
ꢀ
1
Compound
Predicted
class
Observed
class
IC50 (mg mL )
L1210/0
Molt4/C8
CEM/0
1
2
3
4
5
6
1
1
1
1
1
1
1
1
1
1
1
1
66.3
85.2
35.1
103
50.3
98.8
108
98.8
78.1
81.3
97.2
53.2
80.0
138
66.0
82.0
37.1
60.9
characteristic for atoms involved in double bonds [41]. In clas-
sical Markov theory, these numbers are the probabilities with
which the system returns to the initial state. In the present con-
text, they are the probabilities with which electrons return to
M
X
1
u
2
b ¼ d
ð3 Þ
ð4Þ
M
u¼1
where d ¼ 1.5 and M is the number of examples, that
implies repetition of only examples with error values sub-
stantially higher than the mean. It is not very easy to find a
good initial weights set that facilitates the network training.
The union of two of these strategies (RUL and Uniform) of
selection of parameters gave very good results [44].
the atoms at different distances after an arbitrary initial obser-
SR
vation time t [42]. The calculation of pk for any organic or
0
inorganic molecule was carried out using the MARCH-IN-
SIDE software [43].
n
X
SR
k
pk ¼
pii
ð2Þ
i¼1
4
.3. Design and implementation of the neural network
We have a knowledge base of 961 examples, 298 structurally
4
SmartMLP for ANN training
.2. Heuristics added to the platform
heterogeneous anticancer compounds, and 664 non-anticancer
compounds. Active compounds (labelled as 1) are those re-
ported as anticancer compounds by Martin Negwer [45]or Kle-
man et al. [46]. By the contrary, non-anticancer compounds
Artificial neural networks have been used to solve numer-
ous types of problems. Classification is one of the kinds of
problems where they are commonly applied. The backpropa-
gation (BP) algorithm is one of the most popularly applied
to feed-forward training of neural networks, due to its simplic-
ity, its capacity to extract useful information from samples and
to store it implicitly as weights over their connections. How-
ever, this algorithm also has its limitations of practical order,
which are generally accepted and studied by researchers.
Some of these limitations are (1) its convergence towards
a state of minimum error can be extremely slow, principally
if the size of the network is not very big in oppose to the prob-
lem to manipulate, (2) it can stagnate in local minima before
finishing the learning of all samples, (3) it is almost impossible
to select the network design a priori [32e34].
(
labelled as 0) are selected at random from the same database.
Several neural networks were implemented with one and two
hidden layers. We use a combination of strategies for selecting
parameters: RUL and Uniform, as well as the learning coeffi-
cient previously mentioned. Finally the dataset was divided,
choosing a control sample in a random way, equivalent to
2
5% of the total sample. This was repeated several times and
with these datasets different neural networks were built, testing
with different network topologies. Finally, we obtain a neural
network with a 94.04% of accuracy for the training set and
9
1.25% for the external control set [35].
4
.4. Chemistry
Improving the BP algorithm constitutes currently the work
of researchers worldwide. SmartMLP version 1.5 introduces
some improvements to BP training algorithm to enhance its
performance [44]. Learning coefficient (h) as the following
one was able to accelerate the training process.
Melting points were determined using a Reichert Kofler
thermopan or in capillary tubes on a B u¨ chi 510 apparatus
and are uncorrected. IR spectra were recorded on a Perkine
Elmer 1640FT spectrophotometer. H and C NMR spectra
were recorded on a Bruker AMX spectrometer at 300 and
75.47 MHz, respectively, using TMS as internal standard
(chemical shifts in d values, J in hertz). Mass spectra were ob-
1
13
C
N
1
þ N
2
2
þ / þ N
m
h ¼ 2
ꢁ
N þ N þ / þ N
ð3Þ
2
1
2
m
2
tained using
a HewlettePackard 5988A spectrometer.
where C is the number of examples of the training set, m
the number of classes and N is the number of examples
that lie within the ith class. The frequency or order of the
patterns presentation also affects the training of the network.
SmartMLP has implemented three types of selection
patterns:
Elemental analyses were performed using a Perkin Elmer
240B microanalyzer and were within ꢀ0.4 to þ0.4% of calcu-
lated values in all cases. Silica gel (Merck 60, 230e400 mesh)
was used for flash chromatography (FC). Analytical thin layer
chromatography (TLC) was performed on plates precoated
with silica gel (Merck 60 F254, 0.25 mm).
i

584
H. Gonz a´ lez-D ´ı az et al. / European Journal of Medicinal Chemistry 42 (2007) 580e585
4
.4.1. Condensation of pyrimidine bases with
acetic acid (6 mL), and the mixture was heated to reflux (48 h
for compound 4, and 6 h for compound 5). After the reaction
mixture had cooled, the solvent was evaporated and the residue
was purified by FC using 4:1 hexane/ethyl acetate as an eluent.
cyclopentyl bromide. General procedure
A mixture of pyrimidine (1.53 mmol) and a catalytic
amount of ammonium sulphate in hexamethyldisilazane
ꢂ
(
12 mL) was heated at 135 C for 12 h under an Ar atmo-
4
.4.2.1. 5-Chloro-1-cyclopentyluracil (4). Prepared from 1 and
sphere. The resulting clear solution was concentrated in vacuo
under anhydrous conditions to yield the silylated pyrimidine
as a colourless oil. This oil was immediately dissolved in
dry 1,2-dichloroethane (4 mL), and a solution of cyclopentyl
bromide (1.53 mmol) in the same solvent (4 mL) was added.
This mixture was heated to reflux under an inert atmosphere
ꢂ
N-chlorosuccinimide in 68% yield, m.p. 196e198 C. IR: 3167,
3
8
1
031, 2967, 1717, 1661, 1613, 1471, 1316. H NMR (CDCl ):
3
0
.90 (1H, s, NH), 7.45 (1H, s, H-6), 4.90 (1H, m, H-1 ), 2.15
0
0
0
(
3
1
2H, m, 1H-2 þ 1H-5 ), 1.95e1.55 (6H, m, 1H-2 þ H-
0
0
0
13
þ H-4 þ 1H-5 ). C NMR (CDCl ): 159.4, 150.8, 138.5,
3
þ
09.3, 57.8, 31.9, 24.4. MS m/z (%): 216 ([M þ 2] , 10), 214
(
3
48 h for compounds 1 and 2, and one week for compound
) and then cooled, treated with 6:1 methanol/water and fil-
tered. The filtrate was extracted with CH Cl and the solution
þ
þ
þ
(
1
M , 30), 149 ([ClUra þ 2] , 35), 148 ([ClUra þ 1] , 30),
þ
þ
47 ([ClUra] , 100), 146 ([ClUra ꢀ 1] , 72), 103 (25), 69
þ þ
2
2
(
(
[M ꢀ ClUra] , 12), 68 ([M ꢀ ClUra ꢀ 1] , 10), 67
obtained was dried over anhydrous Na SO . The drying agent
2
4
þ
[M ꢀ ClUra ꢀ 2] , 15). Anal. C H ClN O (C, H, N).
was filtered out and the solvent was evaporated in vacuo; the
residue was purified by FC using 3:1 hexane/ethyl acetate as
an eluent.
9
11
2 2
4
.4.2.2. 5-Bromo-1-cyclopentyluracil (5). Prepared from 1 and
ꢂ
N-bromosuccinimide in 93% yield, m.p. 200e202 C. IR:
1
160, 3028, 2965, 2840, 1718, 1654, 1606, 1466, 1272. H
3
NMR (CDCl ): 8.82 (1H, s, NH), 7.52 (1H, s, H-6), 4.90
4
from uracil in 11% yield, m.p. 170e172 C. IR: 3000,
2
.4.1.1. 1-Cyclopentyluracil (1). Compound 1 was prepared
3
ꢂ
0
0
13
0
(
1H, m, H-1 ), 2.14 (2H, m, 1H-2 þ 1H-5 ), 1.90e1.50 (6H,
1
875, 2826, 1678, 1613, 1470, 1421, 1383, 1267. H NMR
0
0
0
0
m, 1H-2 þ H-3 þ H-4 þ 1H-5 ). C NMR (CDCl ): 159.6,
3
(
CDCl ): 8.13 (1H, s, NH), 7.22 (1H, d, H-6, J ¼ 8.1),
3
1
51.1, 141.1, 97.1, 57.9, 31.9, 24.4. MS m/z (%): 260
0
5
2
3
1
.72 (1H, dd, H-5, J ¼ 8.1 and 2.3), 4.90 (1H, m, H-1 ),
þ
þ
þ
(
(
8
[M þ 2] , 36), 258 (M , 36), 193 ([BrUra þ 3] , 100), 192
þ þ þ
0
13
0
0
.12 (2H, m, 1H-2 þ 1H-5 ), 1.90e1.50 (6H, 1H-2 þ H-
[BrUra þ 2] , 91), 191 ([BrUra þ 1] , 98), 190 (BrUra ,
0
0
0
þ H-4 þ 1H-5 ). C NMR (CDCl ): 163.4, 151.3, 141.3,
þ
3
4), 149 (33), 147 (30), 69 ([M ꢀ BrUra] , 13), 68
þ
02.8, 57.1, 31.8, 24.5. MS m/z (%): 180 (M , 19), 113
þ
þ
(
(
[M ꢀ BrUra ꢀ 1] , 11), 67 ([M ꢀ BrUra ꢀ 2] , 16), 53
þ
þ
([Ura þ 1] , 100), 112 (Ura , 16), 69 (11). Anal.
C H N O (C, H, N).
13). Anal. C H BrN O (C, H, N).
2 2
9
11
9
12 2 2
4.4.2.3. 1-Cyclopentyl-5-iodouracil (6). A mixture of 1
4
prepared from 5-methyluracil in 10% yield, m.p. 174e
1
.4.1.2. 1-Cyclopentyl-5-methyluracil (2). Compound 2 was
(93 mg, 0.52 mmol), iodine (260 mg, 1.04 mmol) and
0.75 M nitric acid (0.68 mL) in dioxane (8 mL) was stirred
for 2 h at 100 C. After the reaction mixture had cooled, the
solvent was evaporated in vacuo and the residue was purified
ꢂ
1
76 C. IR: 3174, 3038, 2956, 1690, 1659, 1474, 1269. H
ꢂ
NMR (CDCl ): 9.10 (1H, s, NH), 7.01 (1H, q, H-6, J ¼ 1.1),
3
0
0
0
4
.91 (1H, m, H-1 ), 2.10 (2H, m, 1H-2 þ 1H-5 ), 1.93 (3H,
13
by FC using 4:1 hexane/ethyl acetate as an eluent which gave
ꢂ
0
0
0
d, CH , J ¼ 1.1), 1.83e1.55 (6H, m, 1H-2 þ H-3 þ H-4 þ
6 (148 mg, 94% yield), m.p. 220e222 C. IR: 3148, 1955,
3
1
1839, 1699, 1664, 1599, 1427, 1271. H NMR (CDCl
0
1
3
6
H-5 ). C NMR (CDCl ): 164.1, 151.5, 137.1, 111.3, 56.7,
1.6, 24.5. MS m/z (%): 195 ([M þ 1] , 6), 127 ([Thy þ 1] ,
3
):
9.09 (1H, s, NH), 7.61 (1H, s, H-6), 4.88 (1H, m, H-1 ), 2.13
3
þ
þ
0
þ
0
0
0
0
3), 126 (Thy , 100), 83 (16). Anal. C H N O (C, H, N).
10 14 2 2
(2H, m, 1H-2 þ 1H-5 ), 1.90e1.53 (6H, m, 1H-2 þ H-3 þ
0
0
13
H-4 þ 1H-5 ). C NMR (CDCl ): 160.3, 151.1, 146.2, 68.3,
3
þ
þ
5
4
7.9, 31.9, 24.4. MS m/z (%): 306 (M , 43), 239 ([IUra þ 1] ,
4
.4.1.3. 1-Cyclopentyl-5-fluorouracil (3). Compound 3 was
prepared from 5-fluorouracil in 13% yield, m.p. 178e
þ
8), 238 (IUra , 100), 195 (34). Anal. C H IN O (C, H, N).
9
11
2 2
ꢂ
1
80 C. IR: 3169, 3048, 1710, 1670, 1659, 1261. H NMR
1
(
4
1
CDCl ): 9.45 (1H, s, NH), 7.27 (1H, d, H-6, J ¼ 6.3),
3
0
0
0
.93 (1H, m, H-1 ), 2.10 (2H, m, 1H-2 þ 1H-5 ), 1.85e
13
4.5. Biological activity data and assay
0
0
0
0
.50 (6H, m, 1H-2 þ H-3 þ H-4 þ 1H-5 ).
C NMR
(
1
CDCl ): 157.4 and 157.1 (d, C , J ¼ 27), 150.3 (C ),
3
4
2
All assays were performed in a flat-bottomed 96-well mi-
4
42.6 and 139.5 (d, C , J ¼ 237), 125.8 and 125.3 (d, C ,
5
6
crotiter plates. To each well were added 5 ꢁ 10 L1210/0, or
þ
J ¼ 32), 57.4, 31.7, 24.4. MS m/z (%): 199 ([M þ 1] , 4),
4
7
.5 ꢁ 10 CEM/0, or Molt4/C8 cells and a certain amount of
þ
þ
þ
1
Anal. C H FN O (C, H, N).
98 (M , 25), 131 ([FUra þ 1] , 100), 130 (FUra , 30).
the test compound. The cells were allowed to proliferate for
ꢂ
9
11
2
2
48 h (L1210/0) or 72 h (CEM/0, and Molt4/C8) at 37 C in
a humidified CO controlled atmosphere. At the end of the
2
4
.4.2. Halogenation of 1 using N-halosuccinimide.
General procedure
incubation period, the cells were counted in a Coulter counter
(Coulter electronics Ltd., Harpenden, Herts, England). The
IC50 was defined as the concentration of the compound that
reduced the number of living cells by 50% [47].
To a solution of 1 (100 mg, 0.55 mmol) in acetic acid (3 mL)
was added a solution of N-halosuccinimide (0.60 mmol) in

H. Gonz a´ lez-D ´ı az et al. / European Journal of Medicinal Chemistry 42 (2007) 580e585
585
[20] H. Gonza
lez-D ´ı az, E. Olaz a´ bal, N. Casta n˜ edo, S.I. Hern a´ dez, A. Morales,
´
Acknowledgements
H.S. Serrano, J. Gonz a´ lez, A.R. Ramos de, J. Mol. Model. 8 (2002) 237.
21] H. Gonz a´ lez-D ´ı az, O. Gia, E. Uriarte, I. Hern a´ dez, R. Ramos, M. Chaviano,
S. Seijo, J.A. Castillo, L. Morales, L. Santana, D. Akpaloo, E. Molina,
M. Cruz, L.A. Torres, M.A. Cabrera, J. Mol. Model. 9 (2003) 395.
[22] H. Gonz a´ lez-D ´ı az, S.I. Hern a´ ndez, E. Uriarte, L. Santana, Comput. Biol.
Chem. 27 (2003) 217.
[
The authors thank funds from Xunta de Galicia projects
PXIB20304PR and BTF20302PR as well as from Ministerio
de Sanidad y Consumo project PI061457. The authors are
also indebted to Prof. Dr. Rubio-Gonz a´ lez A. (Post-graduation
and Research, UCLV) and Prof. Dr. Morales S. (VLIR project
coordinator Faculty of Pharmacy, UCLV) by kind attention
and partial assignation of Vlaamse Interuniversitaire Raad
[
23] H. Gonz a´ lez-D ´ı az, A.R. Ramos de, R.R. Molina, Bull. Math. Biol. 65
2003) 991.
24] H. Gonz a´ lez-D ´ı az, A.R. Ramos de, E. Uriarte, Online J. Bioinf.
(2002) 83.
(
[
1
(
VLIR) USO, University Development Cooperation program
[25] A.R. Ramos de, H. Gonz a´ lez-D ´ı az, R. Molina, M.P. Gonz a´ lez, E. Uriarte,
Bioorg. Med. Chem. 12 (2004) 4815.
by partial financial support. Gonz a´ lez-D ´ı az, H. acknowledges
contract as guest-professor from the Department of Organic
Chemistry of the Faculty of Pharmacy at the University of
Santiago de Compostela, Spain in 2006. Bonet, I. acknowl-
edges hospitality and financial support from Department of
Plant Systems Biology, Flanders Interuniversity Institute for
Biotechnology (VIB), Ghent University, Belgium.
[
26] E. Estrada, E. Uriarte, A. Montero, M. Teijeira, L. Santana, E. De Clercq,
J. Med. Chem. 43 (2000) 1975.
[
27] W.A. Denny, The role of medicinal chemistry in the discovery of DNA-
active anticancer drugs: from random searching, through lead develop-
ment, to the novo design, in: M.J. Waring, B.A.J. Ponder (Eds.), The
Search for Anticancer Drugs, Kluwer, Dordrecht, 1992 (chapter 2).
28] H. Kubinyi, J. Recept. Signal Transduct. Res. 19 (1999) 15.
29] W.P. Walters, M.T. Stahl, M.A. Murcko, Drug Discov. Today 3 (1998) 160.
30] K. Ferrante, B. Winograd, R. Canetta, Cancer Chemother. Pharmacol. 43
[
[
[
Appendix. Supplementary material
(
31] E. Menta, M. Palumbo, Expert Opin. Ther. Pat. 8 (1998) 1627.
1999) S61.
[
Supplementary data associated with this article can be found,
in the online version, at doi: 10.1016/j.ejmech.2006.11.016.
[32] I. Bonet-Cruz, Learning Optimization in a MLP Neural Network Applied
to OCR. MICAI 2002: Advances in Artificial Intelligence, LNAI, vol.
2
313, 2002, pp. 292e300.
33] C. Caching, Neural Netw. 7 (1994) 175.
[34] S. Vilar, L. Santana, E. Uriarte, J. Med. Chem. 49 (2006) 1118.
[
References
[
35] H. Van Waterbeemd, Discriminant analysis for activity prediction, in:
R. Manhnhold, Krogsgaard-Larsen, H. Timmerman, H. Van Waterbeemd
[1] H. Kubinyi, J. Taylor, C. Ramdsen, Quantitative Drug Design, in:
C. Hansch (Ed.), Comprehensive Medicinal Chemistry, vol. 4, Pergamon,
(
Eds.), Method and Principles in Medicinal Chemistry, Chemometric
Methods in Molecular Design, vol. 2, VCH, Weinhiem, 1995, pp. 265e282.
36] C. Teran, M. Teijeira, L. Santana, E. Uriarte, A. Casti n˜ eiras, J. Mol.
Struct. 448 (1998) 69.
1
990, pp. 589e643.
[
[
[
[
[
[
[
[
[
[
2] E. Estrada, E. Uriarte, Curr. Med. Chem. 8 (2001) 1573.
3] H. Wiener, J. Am. Chem. Soc. 69 (1947) 17.
37] C. Teran, L. Santana, M. Teijeira, E. Uriarte, J. Balzarini, E. De Clercq,
Pharmazie 53 (1998) 644.
38] H. Gonz a´ lez-D ´ı az, R.R. Molina, E. Uriarte, Bioorg. Med. Chem. Lett. 14
4] J. Devillers, A.T. Balaban, Topological Indices and Related Descriptors
in QSAR and Drug Design, Gordon & Breach, Amsterdam, 2000, pp.
3e41.
(
2004) 4691.
39] H. Gonz a´ lez-D ´ı az, A.R. Ramos de, R.R. Molina, Bioinformatics 19
2003) 2079.
[
5] L.B. Kier, L.H. Hall, Topological Indices and Related Descriptors in
QSAR and QSPR, Gordon and Breach Sci. Pub., Amsterdam, 1999,
pp. 455e489.
6] Y. Marrero-Ponce, H. Gonz a´ lez-D ´ı az, V. Romero-Zaldivar, F. Torrens,
E.A. Castro, Bioorg. Med. Chem. 12 (2004) 5331.
(
40] L. Pauling, The Nature of Chemical Bond, Cornell University Press,
Ithaca, New York, 1939, pp. 2e60.
[
41] E. Eliel, S. Wilen, L. Mander, Stereo Chemistry of Organic Compounds,
John Wiley & Sons Inc., New York, 1994, pp. 103e112.
42] H. Gonz a´ lez-D ´ı az, I. Bastida, N. Casta n˜ edo, O. Nasco, E. Olazabal,
A. Morales, H.S. Serrano, A.R. Ramos de, Bull. Math. Biol. 66 (2004)
[
7] R. Todeschini, V. Consonni, Handbook of Molecular Descriptors, Wiley-
VCH, Weinheim, Germany, 2000.
[
[
8] I. Gutman, V.R. Rosenfield, Theor. Chim. Acta 93 (1996) 191.
9] E. Estrada, Bioinformatics 18 (2002) 1.
1
285.
43] H. Gonz a´ lez-D ´ı az, I. Hern a´ ndez, MARCH-INSIDE 2002, version 1.0
MARkovian CHemicals IN SIlico DEsign). This is a preliminary exper-
[
[
[
[
[
[
10] M.P. Gonz a´ lez, A.H. Morales, R. Molina, Polymer 45 (2004) 2773.
11] A.H. Morales, M.P. Gonz a´ lez, J.B. Rieumont, Polymer 45 (2004) 2045.
12] J.K. Burdett, S. Lee, J. Am. Chem. Soc. 107 (1985) 3063.
13] S. Markovic, I. Gutman, J. Mol. Struct. (Theochem.) 81 (1991) 81.
14] E. Estrada, H. Gonz a´ lez-D ´ı az, J. Chem. Inf. Comput. Sci. 43 (2003) 75.
15] M.P. Gonz a´ lez, H. Gonz a´ lez-D ´ı az, R. Molina, M.A. Cabrera-P e´ rez,
A.R. Ramos de, J. Chem. Inf. Comput. Sci. 43 (2003) 1192.
[
(
imental version. Future professional version shall be available to the pub-
lic. For any information about it send an e-mail to the corresponding
author. <mailto:gonzalezdiazh@yahoo.es>.
[
44] I. Bonet-Cruz, R. Bello, M.M. Garc ´ı a, SmartMLP version 1.5, 2003.
45] M. Negwer, Organic Chemical Drugs and Their Synonyms, Akademie-
Verlag, Berlin, 1987.
[
16] M.A. Cabrera-P e´ rez, H. Gonz a´ lez-D ´ı az, T.C. Fernandez, J.M. Pla-Del-
fina, S.M. Bermejo, Eur. J. Pharm. Biopharm. 53 (2002) 317.
[
17] E. Molina, H. Gonz a´ lez-D ´ı az, M.P. Gonz a´ lez, E. Rodr ´ı guez, E. Uriarte,
J. Chem. Inf. Comput. Sci. 44 (2004) 515.
[
[
46] A. Kleeman, J. Engel, B. Kutscher, D. Reichert, Pharmaceutical Sub-
stances, fourth ed. George Thieme Verlag, Stuttgart, 2001.
[
47] E. De Clercq, in: In vivo and Ex vivo Test Systems to Rationalize
Drug Design and Delivery, Editions Sante, Paris, France, 1994, pp.
[
18] E. Estrada, A. Pe n˜ a, Bioorg. Med. Chem. 8 (2000) 2755.
19] H. Gonz a´ lez-D ´ı az, R. Uriarte, A.R. Ramos de, Bioorg. Med. Chem. 13
[
1
08e125.
(2004) 323.