Combining selectivity and affinity predictions using an integrated Support Vector Machine (SVM) approach: An alternative tool to discriminate between the human adenosine A2A and A3 receptor pyrazolo-triazolo-pyrimidine antagonists bi

Article abstract of DOI:10.1016/j.bmc.2009.05.038

G Protein-coupled receptors (GPCRs) selectivity is an important aspect of drug discovery process, and distinguishing between related receptor subtypes is often the key to therapeutic success. Nowadays, very few valuable computational tools are available f

Full text of DOI:10.1016/j.bmc.2009.05.038

Bioorganic & Medicinal Chemistry 17 (2009) 5259–5274
Contents lists available at ScienceDirect
Bioorganic & Medicinal Chemistry
journal homepage: www.elsevier.com/locate/bmc
Combining selectivity and affinity predictions using an integrated Support
Vector Machine (SVM) approach: An alternative tool to discriminate between
the human adenosine A2A and A3 receptor pyrazolo-triazolo-pyrimidine
antagonists binding sites
Lisa Michielan a, Chiara Bolcato b, Stephanie Federico b, Barbara Cacciari c, Magdalena Bacilieri a,
Karl-Norbert Klotz d, Sonja Kachler d, Giorgia Pastorin e, Riccardo Cardin f, Alessandro Sperduti f,
a,
Giampiero Spalluto b, , Stefano Moro
*
*
a Molecular Modeling Section (MMS), Dipartimento di Scienze Farmaceutiche, Università di Padova, via Marzolo 5, I-35131 Padova, Italy
b Dipartimento di Scienze Farmaceutiche, Università di Trieste, Piazzale Europa 1, I-34127, Trieste, Italy
c Dipartimento di Scienze Farmaceutiche, Università degli Studi di Ferrara, Via Fossato di Mortara 17-19, I-44100 Ferrara, Italy
d Institut für Pharmakologie, Universität of Würzburg, D-97078 Würzburg, Germany
e Department of Pharmacy, National University of Singapore, Block S4, 18 Science Drive 4, Singapore
f Dipartimento di Matematica Pura ed Applicata, Universitá di Padova, Via Trieste 63, I-35121 Padova, Italy
a r t i c l e i n f o
a b s t r a c t
Article history:
Received 19 February 2009
Revised 8 May 2009
Accepted 14 May 2009
Available online 21 May 2009
G Protein-coupled receptors (GPCRs) selectivity is an important aspect of drug discovery process, and
distinguishing between related receptor subtypes is often the key to therapeutic success. Nowadays,
very few valuable computational tools are available for the prediction of receptor subtypes
selectivity.
In the present study, we present an alternative application of the Support Vector Machine (SVM)
and Support Vector Regression (SVR) methodologies to simultaneously describe both A2AR versus A3R
subtypes selectivity profile and the corresponding receptor binding affinities. We have implemented
an integrated application of SVM–SVR approach, based on the use of our recently reported autocor-
related molecular descriptors encoding for the Molecular Electrostatic Potential (autoMEP), to simul-
Keywords:
Adenosine receptors
G protein-coupled receptor
Selectivity profile
Support Vector Machine
QSAR
taneously discriminate A2AR versus A3R antagonists and to predict their binding affinity to the
corresponding receptor subtype of a large dataset of known pyrazolo-triazolo-pyrimidine analogs.
To validate our approach, we have synthetized 51 new pyrazolo-triazolo-pyrimidine derivatives
anticipating both A2AR/A3R subtypes selectivity and receptor binding affinity profiles.
Ó 2009 Published by Elsevier Ltd.
1. Introduction
putational tools are available for the prediction of receptor sub-
types selectivity. Conversely, different in silico approaches are
More than 50% of prescription drugs act on G protein-coupled
receptors (GPCRs). However, some have efficacy problems and lim-
iting side-effects because the compounds do not differentiate be-
tween receptor subtypes. There is thus considerable interest in
attaining therapeutic selectivity by identifying the single receptor
subtype that affects a particular physiology. The goal is to be able
to design drugs without, or at least with less, side-effects, while
retaining the desired function. Nowadays, very few valuable com-
accessible for the prediction of the distinct receptor–ligand affinity.
In particular, the generation of a quantitative structure–activity
relationship (QSAR) is probably one of the most robust and vali-
dated approach to mathematically achieve the correlation between
molecular properties and the corresponding receptor–ligand
affinity.1,2
In the last few years, we have intensively explored the possibil-
ity to discover new potent and selective adenosine receptors (ARs)
antagonists. Briefly, the adenosine receptor (AR) family belongs to
GPCR family A, which includes four different subtypes, referred to
as A1, A2A, A2B and A3, which are widely but differentially distrib-
uted throughout the body.3,4 Interestingly, both A2AR and A3R
receptor selective ligands are becoming increasingly attractive
drugs due to the potential role of this receptor in several physio-
pathological processes.5–9 In particular, A2AR antagonists seem to
Abbreviations: AR, adenosine receptor; h, human; TM, transmembrane; EL2,
second extracellular loop; GPCR, G protein-coupled receptor; SVM, Support Vector
Machine; SVR, Support Vector Regression; autoMEP, autocorrelated Molecular
Electrostatic Potential; QSAR, quantitative structure–activity relationship.
* Corresponding authors. Tel.: +39 049 8275704; fax: +39 049 8275366 (S.M.).
E-mail address: stefano.moro@unipd.it (S. Moro).
0968-0896/$ - see front matter Ó 2009 Published by Elsevier Ltd.
doi:10.1016/j.bmc.2009.05.038

5260
L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
play a role in the reduction of neuronal damage in Parkinson’s or
Huntington’s diseases, while A3R antagonists have a potential
application in the tumor growth inhibition and in the treatment
of glaucoma.4,7,10 Consequently, several receptor-based and li-
gand-based drug design approaches have been carried out with
the aim to improve potency and selectivity of different molecular
scaffolds. Among them, the pyrazolo-triazolo-pyrimidine scaffold
has been used as one of the most versatile scaffold in designing
novelty among adenosine receptors antagonists.11 In particular,
proper substitutions at the N5 and N8 positions have driven the
antagonist selectivity to the human A3R subtype.11 On the other
hand, it has been also demonstrated that the substitution at the po-
sition N7 shifts the selectivity profile to the human A2AR subtype.12
Nevertheless, besides this very empirical rule derived only by
experimental evidences, it is very difficult to anticipate the correct
pharmacological receptor profiles of novel pyrazolo-triazolo-
pyrimidine derivatives and, at the same time, inferring in their
receptor–ligand affinities.
we have decided to select the energetically most stable conformers
produced by the software conformational analysis.
2.3. Training set 1
SVM classification (SVMclass) model: A collection of 104 selective
N7- and N8-substituted pyrazolo-triazolo-pyrimidine analogues
(1–104) has been selected as training set in the SVMclass
model.11,12,32–36 See Supplementary data for details.
2.4. Training set 2
SVR regression (SVR) model: A collection of 104 N8-substituted
pyrazolo-triazolo-pyrimidine derivatives (1–71, 105–137), A2A
and A3 selective and non-selective compounds, has been used as
training set of both human A2A
R and A3R non-linear SVR
models.11,12,32,33,37 See Supplementary data for details.
In the present study, we present an alternative application of
the Support Vector Machine (SVM) and Support Vector Regression
(SVR) methodologies to simultaneously describe both A2AR versus
A3R subtypes selectivity profile and the corresponding receptor
binding affinities. SVM represents a group of supervised learning
techniques, firstly applied in pattern recognition, but now very uti-
lized to solve both classification and regression problems.13,14 In
the last years, several classification problems have been solved,
in fact, using SVM approach, such as, for example, the discrimina-
tion between active and non-active compounds.15–18 Moreover,
Support Vector Regression (SVR) has been widely applied as a
non-linear methodology to derive quantitative structure–activity
relationships for the prediction of different chemical and biological
properties.19–23 Moving from these examples, we have imple-
mented an integrated application of SVM–SVR approach, based
on the use of our recently reported autocorrelated molecular
descriptors encoding for the Molecular Electrostatic Potential
(autoMEP), to simultaneously discriminate A2AR versus A3R antag-
onists and to predict their binding affinity to the corresponding
receptor subtype of a large dataset of known pyrazolo-triazolo-
pyrimidine analogs.24–28 To validate our approach, we have
synthetized 51 new pyrazolo-triazolo-pyrimidine derivatives
anticipating both A2AR/A3R subtypes selectivity and receptor bind-
ing affinity profiles. The statistical quality of both training and val-
idation models are very encouraging.
2.5. Test set
A test set of 51 N8-substituted pyrazolo-triazolo-pyrimidine
analogues (138–188) has been selected to validate both SVMclass
and SVR models. See Table 1.
2.6. Molecular Electrostatic Potential (MEP) calculation
Autocorrelation MEP vectors have been introduced by Gasteiger
and collaborators as molecular descriptors computed on the
molecular surface.38 In the present work, MEPs derive from a clas-
sical point charge model: the electrostatic potential for each mol-
ecule is obtained by moving a unit positive point charge across
the molecular surface, and it is calculated at various points j on this
surface by the following equation:
atoms
X
qi
Vj ¼
rji
i
where qi represents the partial charge of each atom i and rji is the
distance between points j and atom i. Starting from the 3D model
of a molecule and its partial atomic charges, the electrostatic poten-
tial is calculated for points on the molecular surface. Partial atomic
charges were calculated by the PEOE (Partial Equalization of Orbital
Electronegativity) method and its extension to conjugated systems
implemented in ADRIANACODE.
30,39,40 Connolly’s solvent accessible sur-
face with a solvent radius of 2.0 Å has been used to project the cor-
responding MEP. Once the autocorrelation function has been
applied, the autocorrelation vector is derived. Connolly’s solvent
2. Materials and methods
2.1. Computational methodologies
accessible surfaces and the corresponding MEPs have been calcu-
30,39,40
lated by ADRIANACODE
.
All modeling studies were carried out on a linux cluster running
under openMosix architecture.29 Molecular structure building and
autocorrelation MEP descriptors have been carried out using ADRI-
ANACODE software (version 2.0).30 Support Vector Machine classifica-
2.7. Autocorrelation MEP (autoMEP) vectors
The idea of using autocorrelation for the transformation of the
constitution of a molecule into a fixed length representation was
introduced by Moreau and Broto.41,42 They considered that a cer-
tain property p of an atom i can be correlated with the same prop-
erty p of atom j and the products of p values can be summed over
all atom pairs having a certain topological distance d. Each compo-
nent of the autocorrelation vector is consequently calculated as
follows
tion and Support Vector Regression have been performed using
light
SVM
software.31
2.2. Molecular structure building
3D Models of the pyrazolo-triazolo-pyrimidine analogs in the
training set and some analogs in the test set were obtained using
the 3D structure generator Corina, which is an integral part of ADRI-
ANACODE, setting parameters to standard values.30 Conformer selec-
tion is one of the most crucial steps in every 3D-QSAR approach.
Unfortunately, the information about the possible binding mode
of all A2A and A3 receptor antagonists are limited and consequently
X
AðdÞ ¼
pipj
ij
where A is the autocorrelation coefficient referring to atom pairs i,j
at the i,j topological distance d and pi is the atomic property.41,42

L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
5261
Table 1
Biological profile at the hA1 and hA2B receptor subtypes of synthesized (138–188) compounds
O
H
N
R
N
N
N
O
N
N
N
R1
138-188
a
Displacement of specific [3H]-CCPA binding at human A1 receptors expressed in CHO cells (n = 3–6).
b Ki values of the inhibition of NECA-stimulated adenylyl cyclase activity in CHO cells expressing hA2B receptors. Data are expressed as geometric means, with 95% confidence
limits.
Ligands and proteins interact through molecular surfaces and,
therefore, representations of molecular surfaces have to be
sought in the endeavor to understand biological activity. Again,
we are under the restriction of having to represent molecular
surfaces of different size, and consequently Gasteiger and collab-
orators employed the autocorrelation concept to achieve this

5262
L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
n
goal.40,43,44 Starting form the topological autocorrelation exam-
ples of Moreau and Broto, firstly they generated a set of ran-
domly distributed points on the molecular surface, then, all
distances between the surface points were calculated and sorted
into preset intervals dlower ꢀ dupper. The autocorrelation coeffi-
X
1
2
2
min kwk þ C
ni
w;b
i¼1
subject to :
8
i 2 f1; . . . ; ngyiðw ꢁ x þ bÞ P 1 ꢀ ni and ni P 0;
where we recall that y2i fꢀ1;þ1g, w and b are the parameters that control
the function f(x), and the constraints are satisfied with zero error
when it is possible to find a function which is able to ‘classify’ any po-
sitive example (yi = +1) by returning a positive value, that is, f(x) P 1,
and any negative example (yi = ꢀ1) returninga negative value, that is,
f(x) 6 ꢀ1. If such function does not exist, then errors need to be com-
pensated by choosing non-zero values for the corresponding slack
variables ni. The tradeoff between the minimization of the norm of
the weight vector and the empirical error is given by the constant C.
It turns out that the above quadratic constrained minimization
problem can be more easily solved by resorting to the correspond-
ing dual problem
cients are then computed:
X
Aðdlower; dupperÞ ¼ 1=L
pipjðdlower < dij < dupperÞ;
where the i,j distance d belongs to the dlower ꢀ dupper interval and L
is the total number of distances in the same interval. The applica-
tion of this concept made possible to compare different molecular
properties, since this 3D descriptor represents
a compressed
expression of the distribution of the property p on the molecular
surface.
For the calculation of the autocorrelation coefficient we have
applied the default values for parameters computation, since no
improving in statistical model capability was observed by changing
them in a various way. Default parameters values are the follow-
ing: dlower = 1 Å; dupper = 13 Å; L = 12; point density = 10 points/Å2;
vdW radius reduction factor = 1.000. Consequently, we derived 12
autocorrelation vectors per molecule, computed at the 12 (L value)
distances in the interval from 1 to 13 Å with a step width of 1 Å.
We considered that the step width of 1 Å, derived from the parti-
tion in 12 intervals of the global distance 1–13 Å, was sufficient
to describe in an accurate way the distribution of the MEP property
in the molecular surface. The autocorrelation transformation pro-
duces a molecular descriptor which is a unique fingerprint of each
n
n
X
X
1
2
max
ai
ꢀ
aiajyiyjðxi ꢁ xjÞ
a
i¼1
i;j¼1
n
X
subject to :
yiai ¼ 0; and 8i 2 f1; . . . ; ng 0 6 ai 6 C;
i;j¼1
where ai are called the dual variables. The input vectors xi for which
the corresponding dual variables satisfy aoi > 0 are referred to as
support vectors, where ao is the optimal solution vector of the dual
problem. Finally, the decision rule is given by sgn(f(x)).
Non-linearity of the boundary separating positive from negative
samples is achieved by projecting the input vector into a higher
dimensional feature space. In this way the dot product xiꢁxj is re-
placed by a kernel function k(xiꢁxj) representing the dot product
molecule under consideration. Autocorrelation vector have been
30
calculated by ADRIANACODE
.
in the transformed space, that is, k(xiꢁxj) =
U
(xi)ꢁ
U(xj). The decision
2.8. Support Vector Machine (SVM)
function takes the final form
!
n
Support Vector Machines (SVMs) are learning systems origi-
nated from Statistical Learning Theory, proposed by Vapnik.45,46
Here SVMs are applied to solve function approximation problems
(approximation of unknown target functions), by using a training
set represented as pairs of examples, T ¼ fðx1; y1Þ; ðx2; y2Þ; :::;
ðxi; yiÞg, where xi is an input sample and yi is the corresponding de-
sired value, direc2tlfyꢀ1o;þb1gserved from the unknown function. Usually
X
fðxÞ ¼ sgn
a
iyikðxi; xÞ þ b
:
i¼1
Support Vector Regression (SVR):48 If the learning task is defined
as a regression problem, then the mathematical formulation has to
take into account approximation errors. A ‘reasonable’ approxima-
tion is defined by introducing the constraint that for each input xi
xi 2 Rn, while if y
the learning problem is a binary classifica-
we should have j yi ꢀ fðxÞi j6
e, where e is a small positive constant
i
tion problem, if yi 2 R the learning problem is a regression prob-
lem. In both cases, the aim of the learning system is to select
from a set of functions, that is, the hypothesis space, a hypothesis
foptðxÞ that approximates the desired responses yi over the data
set in an optimal fashion, for example, minimizing some risk func-
tional R which weights the cost of the approximation, given by a
loss or cost function that measures the distance between yi and
f(x), with the probability to observe the input–output couple (xi,
yi). In particular, we would like the function f(x) to be a reasonable
estimate of the functional relation between input–output pairs
(prediction or generalization property).
representing the tolerance we allow on approximation errors. This
requirement can be represented by two linear constraints, that is,
ðyi ꢀ w ꢁ xi ꢀ bÞ 6
tolerance are typically linearly penalized by resorting to the linear
-insensitive loss function, defined in the following way
Leðx; y; fÞ ¼ jy ꢀ fðxÞje ¼ maxð0; jy ꢀ fðxÞj ꢀ
Þ:
e
and ðw ꢁ xi þ b ꢀ yiÞ 6
e. Errors above the
e
e
On the basis of the above considerations, the standard SVR
model is defined as
n
X
1
2
2
min kwk þ C
ðni þ nꢂÞ;
i
w;b;ni;nꢂ
The linear SVM is based on: (i) linear hypotheses corresponding
i
P
i¼1
n
to separating hyperplanes, that is, fðxÞ ¼ w ꢁ x þ b ¼ i¼1wixi þ b,
where ꢁ is the dot product between vectors; (ii) the solution of a
quadratic optimization problem that represents a tradeoff between
the minimization of the empirical error, that is, the error over the
training set, and the maximization of the smoothness of f(x).46,47
Non-linear versions of SVM can be obtained by the introduction
of a kernel.47 An example of kernel function is the Gaussian kernel
subject to :
8
i 2 f1; . . . ; ng
ðhw ꢁ xii þ bÞ ꢀ yi 6
e
e
þ ni;
yi ꢀ ðhw ꢁ xii þ bÞ 6
þ nꢂi ;
ni; nꢂ P 0;
i
where a set of slack variables ni; nꢂ, is added to quantify the violation
i
c
kxiꢀxjk2
kðxi; xjÞ ¼ eꢀ
. Let elaborate about these issues, both for clas-
of the imposed constrains: ni account for underestimation of the tar-
get values, while nꢂi account for overestimation of the target values.
Note that, the solution of a linear regression problem results to be a
sification and regression problems, in the following.
Support Vector Classification: Binary classification is widely per-
formed to discriminate a set of compounds in two classes. The
standard formulation for SVM is derived by using the Hinge loss
function and slack variables ni:
tube with radius e
which approximates the data distribution.45–48
Even for SVR a kernel can be used to introduce non-linearity. In
that case, the kernel expansion of the decision function f is

L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
5263
n
X
(d) downfield from tetramethylsilane as internal standard, and J
values are given in hertz. Light petroleum ether refers to the frac-
tions boiling at 40–60 °C. Melting points were determined on a Bu-
chi-Tottoli instrument and are uncorrected. Flash chromatography
was performed using Merck 60–200 mesh silica gel. Elemental
analyses were performed by the microanalytical laboratory of
Dipartimento di Chimica, University of Trieste, and were within
0.4% of the theoretical values for C, H and N. See Supplementary
data for details.
fðx; aꢂi
;
aiÞ ¼
ð
aiꢂ
ꢀ aiÞkðxi; xÞ þ b:
i¼1
The solution of the regression problem is a weighted sum of the ker-
nel function evaluated at the support vectors, defined as the train-
ing points located on the border of the regression tube.
2.9. Classification model validation and evaluation
An extensive n-fold validation procedure was applied to evalu-
ate the prediction capability of the SVM classification model (auto-
MEP/SVMclass). In particular leave-one-out, 10- and 5-fold cross-
validation were performed for model validation. The average, stan-
dard deviation, minimum and maximum rates were collected for
each n-fold cross-validation method. Then, the confusion matrix
was extracted by comparing the experimental and predicted clas-
ses after LOO cross-validation and the test set prediction to confirm
the robustness of our model. In the specific case, hA2AR antagonists
are positive samples, while hA3R antagonists are the negatives. The
percentage (%) of correct predictions, sensitivity and specificity
were calculated from the confusion matrix in the following way:
3.1.1. General procedures for the preparation of 5-[(aryl)carbonyl]
amino-8-(ar)alkyl-2-(2-furyl)-pyrazolo[4,3-e] 1,2,4-triazolo[1,5-c]
pyrimidine (138–188)
Amino compound (45, 47, 51, 57–60, 10 mmol) was dissolved
in freshly distilled THF (15 ml) and appropriate acid chloride
(13 mmol) and triethylamine (13 mmol) were added. The mixture
was refluxed under argon for 18 h. Then the solvent was removed
under reduced pressure and the residue was dissolved in EtOAc
(30 ml) and washed twice with water (15 ml). The organic phase
was dried on Na2SO4 and concentrated under reduced pressure.
The residue was purified by flash chromatography (EtOAc–light
petroleum 4:6) to afford the desired compounds (138–188).
TP þ TN
Correct predictions ð%Þ ¼
;
n
3.1.1.1. 5-[(Diphenylmethyl)carbonyl]amino-8-methyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (139). Yield 85%,
white solid; mp 253 °C (EtOAc–light petroleum); IR (KBr): 3240–
TP
Sensitivity ¼
Specificity ¼
;
TP þ FN
TN
2975, 1680, 1615, 1590, 1515 cmꢀ1 1H NMR (CDCl3) d: 4.01 (s,
;
TN þ FP
3H); 5.62 (s, 1 h); 6.71 (dd, 1H, J = 2, J = 4); 7.04–7.43 (m, 11H);
7.98 (d, 1H, J = 2); 8.79 (s, 1H); 11.38 (br s, 1H). MW 449.16. Anal.
Calcd for C25H19N7O2: C, 66.81; H, 4.26; N, 21.81. Found: C, 66.25;
H, 4.19; N, 22.01.
where TP is the number of true positives, TN is the number of true
negatives, FP are the false positives, FN are the false negatives and n
is the total number of compounds. High percentage (%) values of
correct predictions, sensitivity and specificity correspond to good
modeling performances.
3.1.1.2. 5-[(4-Biphenylmethyl)carbonyl]amino-8-methyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (140). Yield
77%, pale yellow solid; mp 256 °C (EtOAc–light petroleum); IR
(KBr): 3230–2980, 1675, 1620, 1585, 1510 cmꢀ1; 1H NMR (CDCl3)
d: 4.01 (s, 2H); 4.19 (s, 3H); 6.70 (dd, 1H, J = 2, J = 4); 7.21 (d, 1H,
J = 4); 7.25–7.48 (m, 5H); 7.59–7.71 (m, 4H); 7.96 (d, 1H, J = 2);
8.79 (s, 1H); 10.96 (br s, 1H). MW 449.16. Anal. Calcd for
C25H19N7O2: C, 66.81; H, 4.26; N, 21.81. Found: C, 66.43; H, 4.21;
N, 21.77.
3. Chemistry
The desired compounds (138–188) have been prepared by acyl-
ation with the appropriate arylacetyl chloride (1.3 equiv) of the
well-known and previously reported unsubstituted derivatives
45, 47, 51, 57–60 in dry THF at reflux (18 h) in the presence of tri-
ethylamine (1.3 equiv). See Scheme 1. When not commercially
available the acyl chlorides were prepared from the corresponding
acid by treatment with thionyl chloride in the presence of a cata-
lytic amount of dimethyl formamide.
3.1.1.3. 5-[(3-Chlorobenzyl)carbonyl]amino-8-methyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (147). Yield 91%,
yellow solid; mp 212 °C (EtOAc–light petroleum); IR (KBr): 3235–
3.1. General
2975, 1678, 1617, 1588, 1510 cmꢀ1 1H NMR (CDCl3) d: 4.00 (s,
;
2H); 4.17 (s, 3H); 6.73 (dd, 1H, J = 2, J = 4); 7.20 (d, 1H, J = 4); 7.23–
7.38 (m, 3H); 7.43 (s, 1H); 7.99 (d, 1H, J = 2); 8.80 (s, 1H); 11.15 (br
s, 1H). MW 407.81. Anal. Calcd for C19H14N7O2Cl: C, 55.96; H, 3.46;
N, 24.04. Found: C, 55.78; H, 3.39; N, 24.01.
Reactions were routinely monitored by thin-layer chromatogra-
phy (TLC) on silica gel (precoated F254 Merck plates). Infrared spec-
tra (IR) were measured on a Jasco FT-IT instrument. 1H NMR were
determined in CDCl3 or DMSO-d6 solutions with a Varian Gemini
200 spectrometer, peaks positions are given in parts per million
3.1.1.4. 5-[(4-Chlorobenzyl)carbonyl]amino-8-methyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (150). Yield 95%,
brown solid; mp 165 °C (EtOAc–light petroleum); IR (KBr): 3230–
O
2978, 1675, 1614, 1584, 1500 cmꢀ1 1H NMR (CDCl3) d: 4.15 (s,
;
NH2
NH
R
3H); 5.03 (s, 2H); 6.67 (dd, 1H, J = 2, J = 4); 7.01 (d, 2H, J = 9); 7.20
(d, 1H, J = 4); 7.38 (d, 2H, J = 9); 7.97 (d, 1H, J = 2); 8.79 (s, 1H);
11.04 (br s, 1H). MW 407.81. Anal. Calcd for C19H14N7O2Cl: C,
55.96; H, 3.46; N, 24.04. Found: C, 56.12; H, 3.29; N, 24.19.
N
N
N
N
N
i
N
N
O
O
N
N
N
N
R1
N
R1
138-188
3.1.1.5. 5-[(4-Methoxybenzyl)carbonyl]amino-8-methyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (152). Yield
83%, yellow solid; mp 212 °C (EtOAc–light petroleum); IR (KBr):
3227–2983, 1673, 1610, 1582, 1505 cmꢀ1; 1H NMR (CDCl3) d: 3.62
(s, 3H); 3.85 (s, 2H); 4.12 (s, 3H); 6.64 (dd, 1H, J = 2, J = 4); 6.88 (d,
2H, J = 9); 7.19 (d, 1H, J = 4); 7.24 (d, 2H, J = 9); 7.99 (d, 1H, J = 2);
R1 = CH3, 45; R1 = C2H5, 57;
R1 = nC3H7, 58; R1 = nC4H9, 47;
R1 = CH2CH2CH(CH3)2, 51;
R1 = CH2CH2Ph, 59;
R1 = CH2CH2CH2Ph, 60
Scheme 1. Reagents and conditions: (i) RCOCl, TEA, THF, reflux, 18 h.

5264
L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
8.80 (s, 1H); 10.96 (br s, 1H). MW 403.39. Anal. Calcd for C20H17N7O3:
C, 59.55; H, 4.25; N, 24.31. Found: C, 59.47; H, 4.12; N, 24.16.
3.1.1.13. 5-[(2,6-Dichlorobenzyl)carbonyl]amino-8-methyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine(179). Yield
70%, yellow solid; mp 170 °C (EtOAc–light petroleum); IR (KBr):
3228–2983, 1675, 1614, 1591, 1513 cmꢀ1; 1H NMR (CDCl3) d: 4.08
(s, 2H); 4.15 (s, 3H); 6.69 (dd, 1H, J = 2, J = 4); 7.20 (d, 1H, J = 4);
7.31–7.59 (m, 3H); 7.99 (d, 1H, J = 2); 8.84 (s, 1H); 11.15 (br s, 1H).
MW 442.26. Anal. Calcd for C19H13N7O2Cl2: C, 51.60; H, 2.96; N,
22.17. Found: C, 51.32; H, 3.00; N, 22.23.
3.1.1.6. 5-[(2-Thienylmethyl)carbonyl]amino-8-methyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (161). Yield 68%,
brown solid; mp 210 °C (EtOAc–light petroleum); IR (KBr): 3230–
2985, 1670, 1615, 1587, 1515 cmꢀ1 1H NMR (CDCl3) d: 4.16 (s,
;
3H); 4.22 (s, 2H); 6.66 (dd, 1H, J = 2, J = 4); 7.01 (dd, 1H, J = 2, J = 4);
7.07 (d, 1H, J = 2); 7.21 (d, 1H, J = 4); 7.44 (d, 1H, J = 4); 7.97 (d, 1H,
J = 2); 8.81 (s, 1H); 11.03 (br s, 1H). MW 379.40. Anal. Calcd for
C17H13N7O2S: C, 53.82; H, 3.45; N, 25.84. Found: C, 54.03; H, 3.52;
N, 25.88.
3.1.1.14. 5-[(4-Biphenylmethyl)carbonyl]amino-8-ethyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine(151). Yield
72%, pale yellow solid; mp 105 °C (EtOAc–light petroleum); IR (KBr):
3226–2975, 1673, 1610, 1580, 1510 cmꢀ1; 1H NMR (CDCl3) d: 1. 44
(t, 3H, J = 6.7), 4.01 (s, 2H); 4.40 (q, 2H, J = 6.7); 6.72 (dd, 1H, J = 2,
J = 4); 7.19 (d, 1H, J = 4); 7.22–7.58 (m, 5H); 7.62–7.81 (m, 4H);
7.97 (d, 1H, J = 2); 8.79 (s, 1H); 10.96 (br s, 1H). MW 463.49. Anal.
Calcd for C26H21N7O2: C, 67.38; H, 4.57; N, 21.15. Found: C, 66.99;
H, 4.49; N, 21.08.
3.1.1.7. 5-[(3-Thienylmethyl)carbonyl]amino-8-methyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (162). Yield 71%,
brown solid; mp 200 °C (EtOAc–light petroleum); IR (KBr): 3234–
2990, 1668, 1613, 1578, 1512 cmꢀ1; 1H NMR (CDCl3) d: 4.02 (s, 2H);
4.16 (s, 3H); 6.69 (dd, 1H, J = 2, J = 4); 7.07 (d, 1H, J = 4); 7.23 (d, 1H,
J = 4); 7.41 (s, 1H); 7.43 (d, 1H, J = 4); 7.98 (d, 1H, J = 2); 8.80 (s, 1H);
10.99 (br s, 1H). MW 379.40. Anal. Calcd for C17H13N7O2S: C, 53.82;
H, 3.45; N, 25.84. Found: C, 53.66; H, 3.38; N, 25.70.
3.1.1.15. 5-[(2-Thienylmethyl)carbonyl]amino-8-ethyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (160). Yield 65%,
brown solid; mp 113 °C (EtOAc–light petroleum); IR (KBr): 3230–
3.1.1.8. 5-[(b-Naphthylmethyl)carbonyl]amino-8-methyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (163). Yield
87%, yellow solid; mp 209 °C (EtOAc–light petroleum); IR (KBr):
3250–2977, 1682, 1615, 1590, 1500 cmꢀ1; 1H NMR (CDCl3) d: 4.15
(s, 3H); 4.19 (s, 2H); 6.62 (dd, 1H, J = 2, J = 4); 7.20 (d, 1H, J = 4);
7.39–7.58 (m, 3H); 7.68–7.89 (m, 4H); 7.99 (d, 1H, J = 4); 8.79 (s,
1H); 11.01 (br s, 1H). MW 423.43. Anal. Calcd for C23H17N7O2: C,
65.24; H, 4.05; N, 23.16. Found: C, 64.98; H, 4.01; N, 22.99.
2985, 1670, 1618, 1589, 1512 cmꢀ1 1H NMR (CDCl3) d: 1.43 (t, 3H,
;
J = 6.7); 4.21 (s, 2H); 4.43 (q, 2H, J = 6.7); 6.73 (dd, 1H, J = 2, J = 4); 7.00
(dd, 1H, J = 2, J = 4); 7.04 (d, 1H, J = 2); 7.20 (d, 1H, J = 4); 7.41 (d, 1H,
J = 4); 7.99 (d, 1H, J = 2); 8.91 (s, 1H); 11.05 (br s, 1H). MW 393.42. Anal.
Calcd for C18H15N7O2S: C, 54.95; H, 3.84; N, 24.92. Found: C, 55.11; H,
3.91; N, 24.79.
3.1.1.16. 5-[(3-Thienylmethyl)carbonyl]amino-8-ethyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (159). Yield 68%,
hygroscopic brown solid (EtOAc–light petroleum); IR (KBr): 3244–
2991, 1674, 1615, 1590, 1514 cmꢀ1; 1H NMR (CDCl3) d: 1.49 (t, 3H,
J = 6.7); 4.40 (q, 2H, J = 6.7); 4.51 (s, 2H); 6.61 (dd, 1H, J = 2, J = 4);
7.00 (d, 1H, J = 4); 7.15 (d, 1H, J = 4); 7.35 (s, 1H); 7.45 (d, 1H, J = 4);
7.99 (d, 1H, J = 2); 8.82 (s, 1H); 10.35 (br s, 1H). MW 393.42. Anal.
Calcd for C18H15N7O2S: C, 54.95; H, 3.84; N, 24.92. Found: C, 54.68;
H, 3.79; N, 24.77.
3.1.1.9. 5-[(a-Naphthylmethyl)carbonyl]amino-8-methyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (164). Yield
87%, pale yellow solid; mp 205 °C (EtOAc–light petroleum); IR (KBr):
3240–2985, 1675, 1613, 1584, 1511 cmꢀ1; 1H NMR(CDCl3)d: 4.13(s,
3H); 4.42 (s, 2H); 6.63 (dd, 1H, J = 2, J = 4); 7.19 (d, 1H, J = 4); 7.40–
7.61 (m, 4H); 7.75–7.85 (m, 3H); 7.98 (d, 1H, J = 4); 8.78 (s, 1H);
11.08 (br s, 1H). MW 423.43. Anal. Calcd for C23H17N7O2: C, 65.24;
H, 4.05; N, 23.16. Found: C, 65.38; H, 4.04; N, 23.22.
3.1.1.17. 5-[(b-Naphthylmethyl)carbonyl]amino-8-ethyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (158). Yield 91%, pale
yellow solid; mp 97 °C (EtOAc–light petroleum); IR (KBr): 3235–2987,
3.1.1.10. 5-[(4-Trifluorobenzyl)carbonyl]amino-8-methyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (168). Yield
97%, pale yellow solid; mp 212 °C (EtOAc–light petroleum); IR
1676, 1620, 1587, 1505 cmꢀ1 1H NMR (CDCl3) d: 1.47 (t, 3H, J = 6.7);
;
(KBr): 3225–2980, 1670, 1613, 1580, 1505 cmꢀ1
;
1H NMR (CDCl3)
4.18 (s, 2H); 4.41 (q, 2H, J = 6.7); 6.78 (dd, 1H, J = 2, J = 4); 7.21 (d, 1H,
J = 4); 7.36–7.60 (m, 3H); 7.77–7.99 (m, 4H); 8.01 (d, 1H, J = 4); 8.83 (s,
1H); 11.05 (br s, 1H). MW 437.45. Anal. Calcd for C24H19N7O2: C, 65.89;
H, 4.38; N, 22.41. Found: C, 65.66; H, 4.37; N, 22.29.
d: 4.08 (s, 2H); 4.12 (s, 3H); 6.68 (dd, 1H, J = 2, J = 4); 7.21 (d, 1H,
J = 4); 7.61 (d, 2H, J = 9); 7.74 (d, 2H, J = 9); 7.97 (d, 1H, J = 2); 8.80
(s, 1H); 11.19 (br s, 1H). MW 441.37. Anal. Calcd for C20H14N7O2F3:
C, 54.43; H, 3.20; N, 22.21. Found: C, 54.71; H, 3.28; N, 22.38.
3.1.1.18. 5-[(a-Naphthylmethyl)carbonyl]amino-8-ethyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (157). Yield 90%, pale
3.1.1.11. 5-[(4-Chlorophenoxymethyl)carbonyl]amino-8-methyl-
2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (171). Yield
83%, brown solid; mp 145 °C (EtOAc–light petroleum); IR (KBr):
yellow solid; mp 135 °C (EtOAc–light petroleum); IR (KBr): 3238–
2980, 1676, 1611, 1580, 1500 cmꢀ1 1H NMR (CDCl3) d: 1.46 (t, 3H,
;
3230–2980, 1670, 1615, 1590, 1515 cmꢀ1
;
1H NMR (CDCl3) d: 4.28
J = 6.7); 4.41 (q, 2H, J = 6.7); 4.53 (s, 2H); 6.67 (dd, 1H, J = 2, J = 4); 7.20
(d, 1H, J = 4); 7.41–7.64 (m, 4H); 7.79–7.99 (m, 3H); 8.01 (d, 1H, J = 4);
8.84 (s, 1H); 11.05 (br s, 1H). MW 437.45. Anal. Calcd for C24H19N7O2:
C, 65.89; H, 4.38; N, 22.41. Found: C, 66.02; H, 4.42; N, 22.53.
(s, 3H); 5.03 (s, 2H); 6.63 (dd, 1H, J = 2, J = 4); 7.01 (d, 2H, J = 9);
7.18 (d, 1H, J = 4); 7.31 (d, 2H, J = 9); 7.98 (d, 1H, J = 2); 8.78 (s,
1H); 11.03 (br s, 1H). MW 423.81. Anal. Calcd for C19H14N7O3Cl: C,
53.85; H, 3.33; N, 23.13. Found: C, 54.02; H, 3.37; N, 23.19.
3.1.1.19. 5-[(Diphenylmethyl)carbonyl]amino-8-n-propyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (148). Yield
85%, white solid; mp 212 °C (EtOAc–light petroleum); IR (KBr):
3235–2980, 1673, 1610, 1592, 1510 cmꢀ1; 1H NMR (CDCl3) d: 0.91
(t, 3H, J = 6.7), 1.81–1.99 (m, 2H); 4.39 (t, 2H, J = 6.7); 5.61 (s, 1H);
6.67 (dd, 1H, J = 2, J = 4); 7.08–7.58 (m, 11H); 7.99 (d, 1H, J = 2);
8.81 (s, 1H); 11.35 (br s, 1H). MW 477.52. Anal. Calcd for
C27H23N7O2: C, 67.91; H, 4.85; N, 20.53. Found: C, 67.69; H, 4.77;
N, 20.42.
3.1.1.12. 5-[(4-Fluorobenzyl)carbonyl]amino-8-methyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (177). Yield 86%,
brown solid; mp 192 °C (EtOAc–light petroleum); IR (KBr): 3232–
2986, 1673, 1616, 1590, 1511 cmꢀ1 1H NMR (CDCl3) d: 4.01 (s,
;
2H); 4.18 (s, 3H); 6.66 (dd, 1H, J = 2, J = 4); 7.21 (d, 1H, J = 4); 7.23
(d, 2H, J = 9); 7.41 (d, 2H, J = 9); 7.98 (d, 1H, J = 2); 8.81 (s, 1H);
11.00 (br s, 1H). MW 391.36. Anal. Calcd for C19H14N7O2F: C, 58.31;
H, 3.61; N, 25.05. Found: C, 58.07; H, 3.62; N, 24.96.

L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
5265
3.1.1.20. 5-[(4-Biphenylmethyl)carbonyl]amino-8-n-propyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (149). Yield
79%, pale yellow solid; mp 218 °C (EtOAc–light petroleum); IR
4.39 (t, 2H, J = 6.7); 6.75 (dd, 1H, J = 2, J = 4); 7.08 (d, 1H, J = 4); 7.21
(d, 1H, J = 4); 7.41 (s, 1H); 7.58 (d, 1H, J = 4); 7.99 (d, 1H, J = 2); 8.82
(s, 1H); 10.97 (br s, 1H). MW 421.48. Anal. Calcd for C20H19N7O2S:
C, 56.99; H, 4.54; N, 23.26. Found: C, 56.69; H, 4.47; N, 23.19.
(KBr): 3235–2985, 1676, 1620, 1587, 1500 cmꢀ1 1H NMR (CDCl3)
;
d: 0.93 (t, 3H, J = 6.7), 1.83–1.94 (m, 2H); 4.01 (s, 2H); 4.37 (t, 2H,
J = 6.7); 6.76 (dd, 1H, J = 2, J = 4); 7.20 (d, 1H, J = 4); 7.25–7.59 (m,
5H); 7.61–7.79 (m, 4H); 7.98 (d, 1H, J = 2); 8.81 (s, 1H); 11.03 (br s,
1H). MW 477.52. Anal. Calcd for C27H23N7O2: C, 67.91; H, 4.85; N,
20.53. Found: C, 68.01; H, 4.86; N, 20.57.
3.1.1.27. 5-[(a-Naphthylmethyl)carbonyl]amino-8-n-butyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine(167). Yield
75%, pale yellow solid; mp 95 °C (EtOAc–light petroleum); IR (KBr):
3227–2982, 1673, 1619, 1580, 1500 cmꢀ1; 1H NMR (CDCl3) d: 0.85 (t,
3H, J = 6.7), 1.14–1.29 (m, 2H); 1.73–1.98 (m, 2H); 4.35 (t, 2H,
J = 6.7); 4.49 (s, 2H); 6.73 (dd, 1H, J = 2, J = 4); 7.22 (d, 1H, J = 4);
7.39–7.61 (m, 4H); 7.78–8.00 (m, 3H); 8.01 (d, 1H, J = 4); 8.82 (s,
1H); 11.04 (br s, 1H). MW 465.51. Anal. Calcd for C26H23N7O2: C,
67.08; H, 4.98; N, 21.06. Found: C, 67.23; H, 4.95; N, 21.01.
3.1.1.21. 5-[(b-Naphthylmethyl)carbonyl]amino-8-n-propyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (166). Yield
82%, pale yellow solid; mp 80 °C (EtOAc–light petroleum); IR
(KBr): 3230–2980, 1680, 1615, 1584, 1515 cmꢀ1 1H NMR (CDCl3)
;
d: 0.85 (t, 3H, J = 6.7), 1.80–2.01 (m, 2H); 4.19 (s, 2H); 4.35 (t, 2H,
J = 6.7); 6.75 (dd, 1H, J = 2, J = 4); 7.23 (d, 1H, J = 4); 7.40–7.62 (m,
3H); 7.80–7.95 (m, 4H); 7.99 (d, 1H, J = 4); 8.81 (s, 1H); 11.07 (br s,
1H). MW 451.48. Anal. Calcd for C25H21N7O2: C, 66.51; H, 4.69; N,
21.72. Found: C, 66.23; H, 4.60; N, 21.80.
3.1.1.28. 5-[(Diphenylmethyl)carbonyl]amino-8-isopentyl-2-
(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimi-
dine (141). Yield 78%, white solid; mp 180 °C (EtOAc–light petro-
leum); IR (KBr): 3235–2980, 1677, 1618, 1585, 1505 cmꢀ1; 1H NMR
(CDCl3) d: 0.83 (d, 6H, J = 7); 1.35–1.45 (m, 1H); 1.71–1.81 (m, 2H);
4.38 (t, 2H. J = 6.7); 5.64 (s, 1H); 6.75 (dd, 1H, J = 2, J = 4); 7.14–7.45
(m, 11H); 7.97 (d, 1H, J = 2); 8.83 (s, 1H); 11.35 (br s, 1H). MW
465.51. Anal. Calcd for C29H27N7O2: C, 68.89; H, 5.38; N, 19.39.
Found: C, 68.57; H, 5.23; N, 19.48.
3.1.1.22. 5-[(a-Naphthylmethyl)carbonyl]amino-8-n-propyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (165). Yield
75%, pale yellow solid; mp 170 °C (EtOAc–light petroleum); IR
(KBr): 3230–2985, 1675, 1615, 1583, 1506 cmꢀ1 1H NMR (CDCl3)
;
d: 0.85 (t, 3H, J = 6.7), 1.79–1.99 (m, 2H); 4.35 (t, 2H, J = 6.7); 4.53
(s, 2H); 6.77 (dd, 1H, J = 2, J = 4); 7.22 (d, 1H, J = 4); 7.40–7.63 (m,
4H); 7.79–8.00 (m, 3H); 8.03 (d, 1H, J = 4); 8.83 (s, 1H); 11.18 (br s,
1H). MW 451.48. Anal. Calcd for C25H21N7O2: C, 66.51; H, 4.69; N,
21.72. Found: C, 66.30; H, 4.73; N, 21.67.
3.1.1.29. 5-[(4-Biphenylmethyl)carbonyl]amino-8-isopentyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (142). Yield
74%, pale yellow solid; mp 130 °C (EtOAc–light petroleum); IR
(KBr): 3225–2975, 1672, 1610, 1575, 1513 cmꢀ1 1H NMR (CDCl3)
;
d: 0.84 (d, 6H, J = 7); 1.38–1.57 (m, 1H); 1.70–1.82 (m, 2H); 4.01 (s,
2H); 4.40 (t, 2H. J = 6.7); 6.72 (dd, 1H, J = 2, J = 4); 7.21 (d, 1H, J = 4);
7.26–7.51 (m, 5H); 7.57–7.69 (m, 4H); 7.97 (d, 1H, J = 2); 8.82 (s,
1H); 11.03 (br s, 1H). MW 465.51. Anal. Calcd for C29H27N7O2: C,
68.89; H, 5.38; N, 19.39. Found: C, 68.63; H, 5.31; N, 19.43.
3.1.1.23. 5-[(Diphenylmethyl)carbonyl]amino-8-n-butyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (145). Yield
75%, white solid; mp 195 °C (EtOAc–light petroleum); IR (KBr):
3230–2985, 1672, 1615, 1600, 1520 cmꢀ1; 1H NMR (CDCl3) d: 0.81
(t, 3H, J = 6.7), 1.01–1.28 (m, 2H); 1.78–1.99 (m, 2H); 4.38 (t, 2H,
J = 6.7); 5.65 (s, 1H); 6.68 (dd, 1H, J = 2, J = 4); 7.04–7.55 (m, 11H);
7.98 (d, 1H, J = 2); 8.82 (s, 1H); 11.37 (br s, 1H). MW 491.54. Anal.
Calcd for C28H25N7O2: C, 68.42; H, 5.13; N, 19.95. Found: C, 68.17;
H, 5.08; N, 20.02.
3.1.1.30. 5-[(3-Chlorobenzyl)carbonyl]amino-8-isopentyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (172). Yield
89%, pale yellow solid; mp 95 °C (EtOAc–light petroleum); IR
(KBr): 3235–2987, 1676, 1610, 1580, 1500 cmꢀ1 1H NMR (CDCl3)
;
d: 0.83 (d, 6H, J = 7); 1.35–1.55 (m, 1H); 1.69–1.80 (m, 2H); 4.04 (s,
2H); 4.40 (t, 2H. J = 6.7); 6.69 (dd, 1H, J = 2, J = 4); 7.19 (d, 1H,
J = 4); 7.22–7.34 (m, 3H); 7.44 (s, 1H); 8.00 (d, 1H, J = 2); 8.78 (s,
1H); 11.05 (br s, 1H). MW 463.92. Anal. Calcd for C23H22N7O2Cl: C,
59.55; H, 4.78; N, 21.13. Found: C, 59.27; H, 4.59; N, 21.10.
3.1.1.24. 5-[(4-Biphenylmethyl)carbonyl]amino-8-n-butyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (146). Yield
83%, pale yellow solid; mp 180 °C (EtOAc–light petroleum); IR (KBr):
3235–2985, 1676, 1620, 1587, 1500 cmꢀ1; 1H NMR (CDCl3) d: 0.83 (t,
3H, J = 6.7), 1.05–1.25 (m, 2H); 1.69–1.99 (m, 2H); 4.00 (s, 2H); 4.39
(t, 2H, J = 6.7); 6.63 (dd, 1H, J = 2, J = 4); 7.18 (d, 1H, J = 4); 7.21–7.55
(m, 5H); 7.59–7.76 (m, 4H); 7.93 (d, 1H, J = 2); 8.81 (s, 1H); 11.00 (br
s, 1H). MW 491.54. Anal. Calcd for C28H25N7O2: C, 68.42; H, 5.13; N,
19.95. Found: C, 68.63; H, 5.13; N, 20.05.
3.1.1.31. 5-[(4-Chlorobenzyl)carbonyl]amino-8-isopentyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (173). Yield
86%, white solid; mp 92 °C (EtOAc–light petroleum); IR (KBr):
3225–2975, 1672, 1610, 1590, 1507 cmꢀ1 1H NMR (CDCl3) d:
;
0.85 (d, 6H, J = 7); 1.36–1.55 (m, 1H); 1.66–1.81 (m, 2H); 4.38
(t, 2H. J = 6.7); 5.01 (s, 2H); 6.68 (dd, 1H, J = 2, J = 4); 7.01 (d,
2H, J = 9); 7.19 (d, 1H, J = 4); 7.37 (d, 2H, J = 9); 7.99 (d, 1H,
J = 2); 8.90 (s, 1H); 11.15 (br s, 1H). MW 463.92. Anal. Calcd
for C23H22N7O2Cl: C, 59.55; H, 4.78; N, 21.13. Found: C, 59.83;
H, 4.81; N, 21.09.
3.1.1.25. 5-[(2-Thienylmethyl)carbonyl]amino-8-n-butyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (169). Yield
66%, hygroscopic brown solid; (EtOAc–light petroleum); IR (KBr):
3230–2975, 1675, 1610, 1580, 1515 cmꢀ1 1H NMR (CDCl3) d: 0.81
;
(t, 3H, J = 6.7), 1.08–1.35 (m, 2H); 1.68–2.00 (m, 2H); 4.21 (s, 2H);
4.37 (t, 2H, J = 6.7); 6.66 (dd, 1H, J = 2, J = 4); 6.99 (dd, 1H, J = 2,
J = 4); 7.02 (d, 1H, J = 2); 7.22 (d, 1H, J = 4); 7.41 (d, 1H, J = 4); 7.98
(d, 1H, J = 2); 8.83 (s, 1H); 11.03 (br s, 1H). MW 421.48. Anal. Calcd
for C20H19N7O2S: C, 56.99; H, 4.54; N, 23.26. Found: C, 57.19; H,
4.49; N, 23.31.
3.1.1.32. 5-[(b-Naphthylmethyl)carbonyl]amino-8-isopentyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (153). Yield
80%, yellow solid; mp 180 °C (EtOAc–light petroleum); IR
(KBr): 3225–2970, 1676, 1610, 1584, 1505 cmꢀ1
;
1H NMR
(CDCl3) d: 0.84 (d, 6H, J = 7); 1.30–1.57 (m, 1H); 1.69–1.81 (m,
2H); 4.21 (s, 2H); 4.41 (t, 2H. J = 6.7); 6.68 (dd, 1H, J = 2, J = 4);
7.19 (d, 1H, J = 4); 7.36–7.55 (m, 3H); 7.67–7.89 (m, 4H); 7.98
(d, 1H, J = 4); 8.81 (s, 1H); 11.08 (br s, 1H). MW 479.53. Anal.
Calcd for C27H25N7O2: C, 67.63; H, 5.25; N, 20.45. Found: C,
67.30; H, 5.18; N, 20.29.
3.1.1.26. 5-[(3-Thienylmethyl)carbonyl]amino-8-n-butyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (170). Yield
60%, hygroscopic brown solid (EtOAc–light petroleum); IR (KBr):
3230–2985, 1673, 1615, 1580, 1512 cmꢀ1; 1H NMR (CDCl3) d: 0.83
(t, 3H, J = 6.7), 1.04–1.38 (m, 2H); 1.69–2.01 (m, 2H); 4.01 (s, 2H);

5266
L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
3.1.1.33. 5-[(4-Trifluorobenzyl)carbonyl]amino-8-isopentyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (175). Yield
97%, pale yellow solid; mp 95 °C (EtOAc–light petroleum); IR
MW 469.52. Anal. Calcd for C24H19N7O2S: C, 61.39; H, 4.08; N, 20.88.
Found: C, 61.08; H, 4.13; N, 20.65.
(KBr): 3230–2984, 1684, 1620, 1575, 1512 cmꢀ1
;
1H NMR (CDCl3)
3.1.1.40. 5-[(3-Thienylmethyl)carbonyl]amino-8-(2-phenyl
ethyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrim-
idine (156). Yield 67%, hygroscopic brown solid; (EtOAc–light
petroleum); IR (KBr): 3220–2980, 1675, 1610, 1580,
d: 0.86 (d, 6H, J = 7); 1.38–1.57 (m, 1H); 1.67–1.83 (m, 2H); 4.13 (s,
2H); 4.39 (t, 2H. J = 6.7); 6.66 (dd, 1H, J = 2, J = 4); 7.20 (d, 1H,
J = 4); 7.61 (d, 2H, J = 9); 7.72 (d, 2H, J = 9); 7.99 (d, 1H, J = 2); 8.82
(s, 1H); 11.01 (br s, 1H). MW 497.47. Anal. Calcd for C24H22N7O2F3:
C, 57.94; H, 4.46; N, 19.71. Found: C, 58.07; H, 4.45; N, 19.84.
1512 cmꢀ1 1H NMR (CDCl3) d: 3.46 (t, 2H, J = 7); 4.45 (s, 2H);
;
4.52 (t, 2H, J = 7); 6.66 (dd, 1H, J = 2, J = 4); 7.01 7.39 (m, 9H);
7.98 (d, 1H, J = 2); 8.79 (s, 1H); 10.87 (br s, 1H). MW 469.52.
Anal. Calcd for C24H19N7O2S: C, 61.39; H, 4.08; N, 20.88. Found:
C, 61.65; H, 4.01; N, 20.57.
3.1.1.34. 5-[(4-Fluorobenzyl)carbonyl]amino-8-methyl-2-(2-furyl)-
pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (176). Yield 90%,
pale yellow solid; mp 133 °C (EtOAc–light petroleum); IR (KBr):
3235–2990, 1680, 1610, 1585, 1510 cmꢀ1; 1H NMR (CDCl3) d: 0.82
(d, 6H, J = 7); 1.31–1.54 (m, 1H); 1.66–1.83 (m, 2H); 3.99 (s, 2H);
4.39 (t, 2H. J = 6.7); 6.68 (dd, 1H, J = 2, J = 4); 7.14 (d, 1H, J = 4); 7.19
(d, 2H, J = 9); 7.37 (d, 2H, J = 9); 7.95 (d, 1H, J = 2); 8.86 (s, 1H);
11.02 (br s, 1H). MW 447.46. Anal. Calcd for C23H22N7O2F: C, 61.74;
H, 4.96; N, 21.91. Found: C, 61.96; H, 4.79; N, 22.05.
3.1.1.41. 5-[(b-Naphthylmethyl)carbonyl]amino-8-(2-phenyl-
ethyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine
(154). Yield 77%, white solid; mp 113 °C (EtOAc–light petroleum);
IR (KBr): 3220–2970, 1681, 1610, 1575, 1510 cmꢀ1 1H NMR
;
(CDCl3) d: 3.43 (t, 2H, J = 7); 4.40 (s, 2H); 4.61 (t, 2H, J = 7 Hz);
6.66 (dd, 1H, J = 2 Hz, J = 4 Hz); 7.07–7.14 (m, 5H); 7.23–7.36 (m,
4H); 7.83–7.99 (m, 4H); 8.02 (d, 1H, J = 4 Hz); 8.84 (s, 1H); 11.11
(br s, 1H). MW 513.55. Anal. Calcd for C30H23N7O2: C, 70.16; H,
4.51; N, 19.09. Found: C, 69.88; H, 4.43; N, 19.25.
3.1.1.35. 5-[(2,6-Dichlorobenzyl)carbonyl]amino-8-isopentyl-2-(2-
furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (178). Yield
70%, white solid; mp 130 °C (EtOAc–light petroleum); IR (KBr):
3225–2975, 1677, 1615, 1580, 1506 cmꢀ1; 1H NMR (CDCl3) d: 0.86
(d, 6H, J = 7); 1.33–1.56 (m, 1H); 1.71–1.85 (m, 2H); 4.37 (s, 2H);
4.39 (t, 2H. J = 6.7); 6.67 (dd, 1H, J = 2, J = 4); 7.19 (d, 1H, J = 4);
7.29–7.57 (m, 3H); 8.01 (d, 1H, J = 2); 8.80 (s, 1H); 11.03 (br s, 1H).
MW 447.46. Anal. Calcd for C23H21N7O2Cl2: C, 55.10; H, 4.82; N,
19.56. Found: C, 54.95; H, 4.77; N, 19.63.
3.1.1.42. 5-[(Diphenylmethyl)carbonyl]amino-8-(3-phenylpro-
pyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine
(181). Yield 80%, pale yellow solid; mp 193 °C (EtOAc–light petro-
leum); IR (KBr): 3225–2980, 1677, 1610, 1595, 1515 cmꢀ1 1H
;
NMR (CDCl3) d: 2.20 (t, 2H, J = 7); 2.55–2.65 (m, 2H); 4.41 (t, 2H,
J = 7); 5.68 (s, 1H); 6.75 (dd, 1H, J = 2, J = 4); 7.01–7.57 (m, 16H);
7.99 (d, 1H, J = 2); 8.87 (s, 1H); 11.39 (br s, 1H). MW 553.61. Anal.
Calcd for C33H27N7O2: C, 71.59; H, 4.92; N, 17.71. Found: C, 71.63;
H, 5.01; N, 17.93.
3.1.1.36. 5-[(Diphenylmethyl)carbonyl]amino-8-(2-phenylethyl)-
2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine (143). Yield
94%, white solid; mp 109 °C (EtOAc–light petroleum); IR (KBr): 3215–
2970, 1680, 1615, 1580, 1510 cmꢀ1 1H NMR (CDCl3) d: 3.46 (t, 2H,
;
3.1.1.43. 5-[(4-Biphenylmethyl)carbonyl]amino-8-(3-phenyl-
propyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine
(188). Yield 76%, yellow solid; mp 150 °C (EtOAc–light petro-
J = 7); 4.52 (t, 2H, J = 7); 5.60 (s, 1H); 6.71 (dd, 1H, J = 2, J = 4); 7.05–
7.48 (m, 16H); 7.98 (d, 1H, J = 2); 8.80 (s, 1H); 11.28 (br s, 1H). MW
539.59. Anal. Calcd for C32H25N7O2: C, 71.23; H, 4.67; N, 18.17. Found:
C, 70.95; H, 4.72; N, 18.21.
leum); IR (KBr): 3225–2980, 1674, 1610, 1575, 1515 cmꢀ1 1H
;
NMR (CDCl3) d: 2.20 (t, 2H, J = 7); 2.55–2.65 (m, 2H); 4.41 (t, 2H,
J = 7); 4.59 (s, 2H); 6.61 (dd, 1H, J = 2, J = 4); 7.00–7.73 (m, 15H);
8.17 (d, 1H, J = 2); 9.13 (s, 1H); 11.33 (br s, 1H). MW 553.61. Anal.
Calcd for C33H27N7O2: C, 71.59; H, 4.92; N, 17.71. Found: C, 71.21;
H, 4.86; N, 17.69.
3.1.1.37. 5-[(4-Biphenylmethyl)carbonyl]amino-8-(2-phenyl-
ethyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine
(144). Yield 80%, pale yellow solid; mp 95 °C (EtOAc–light petro-
leum); IR (KBr): 3225–2975, 1678, 1615, 1580, 1510 cmꢀ1 1H
;
3.1.1.44. 5-[(3-Chlorobenzyl)carbonyl]amino-8-(3-phenylpro-
pyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine
(184). Yield 95%, white solid; mp 98 °C (EtOAc–light petroleum); IR
(KBr): 3230–2985, 1675, 1610, 1580, 1508 cmꢀ1; 1H NMR (CDCl3) d:
2.20 (t, 2H, J = 7); 2.55–2.65 (m, 2H); 4.01 (s, 2H); 4.41 (t, 2H, J = 7);
6.73 (dd, 1H, J = 2, J = 4); 7.02–7.61 (m, 10H); 7.96 (d, 1H, J = 2); 8.86
(s, 1H); 11.09 (br s, 1H). MW 511.96. Anal. Calcd for C27H22N7O2Cl: C,
63.34; H, 4.33; N, 19.15. Found: C, 63.07; H, 4.22; N, 19.01.
NMR (CDCl3) d: 3.46 (t, 2H, J = 7); 4.01 (s, 2H); 4.52 (t, 2H, J = 7);
6.70 (dd, 1H, J = 2, J = 4); 7.06–7.69 (m, 15H); 7.96 (d, 1H, J = 2);
8.79 (s, 1H); 11.01 (br s, 1H). MW 539.59. Anal. Calcd for
C32H25N7O2: C, 71.23; H, 4.67; N, 18.17. Found: C, 71.13; H, 4.55;
N, 18.08.
3.1.1.38. 5-[(3-Chlorobenzyl)carbonyl]amino-8-(2-phenyl-
ethyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrim-
idine (174). Yield 89%, pale yellow solid; mp 83 °C (EtOAc–
light petroleum); IR (KBr): 3235–2990, 1679, 1613, 1577,
3.1.1.45. 5-[(4-Chlorobenzyl)carbonyl]amino-8-(3-phenylpro-
pyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine
(185). Yield 78%, yellow solid; mp 101 °C (EtOAc–light petro-
1505 cmꢀ1 1H NMR (CDCl3) d: 3.48 (t, 2H, J = 7); 5.09 (s, 2H);
;
4.50 (t, 2H, J = 7); 6.78 (dd, 1H, J = 2, J = 4); 7.19 (d, 1H, J = 4);
7.21–7.40 (m, 9H); 7.99 (d, 1H, J = 2); 8.68 (s, 1H); 10.99 (br
s, 1H). MW 497.94. Anal. Calcd for C26H20N7O2Cl: C, 62.71; H,
4.05; N, 19.69. Found: C, 62.34; H, 3.98; N, 19.47.
leum); IR (KBr): 3225–2980, 1674, 1610, 1580, 1508 cmꢀ1 1H
;
NMR (CDCl3) d: 2.22 (t, 2H, J = 7); 2.52–2.64 (m, 2H); 4.40 (t, 2H,
J = 7); 4.53 (s, 2H); 6.60 (dd, 1H, J = 2, J = 4); 7.04–7.58 (m, 10H);
8.18 (d, 1H, J = 2); 9.10 (s, 1H); 11.05 (br s, 1H). MW 511.96. Anal.
Calcd for C27H22N7O2Cl: C, 63.34; H, 4.33; N, 19.15. Found: C,
63.51; H, 4.34; N, 19.22.
3.1.1.39. 5-[(2-Thienylmethyl)carbonyl]amino-8-(2-phenylethyl)-2-
(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c] pyrimidine (155). Yield
71%, hygroscopic brown solid; (EtOAc–light petroleum); IR (KBr):
3.1.1.46. 5-[(b-Naphthylmethyl)carbonyl]amino-8-(3-phenyl-
propyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrim-
idine (182). Yield 77%, white solid; mp 113 °C (EtOAc–light
3220–2980, 1678, 1610, 1585, 1510 cmꢀ1 1H NMR (CDCl3) d: 3.43 (t,
;
2H, J = 7); 4.54 (t, 2H, J = 7); 4.71 (s, 2H); 6.69 (dd, 1H, J = 2, J = 4);
7.02–7.41 (m, 9H); 7.97 (d, 1H, J = 2); 8.78 (s, 1H); 10.11 (br s, 1H).
petroleum); IR (KBr): 3225–2980, 1679, 1610, 1580, 1510 cmꢀ1
;

L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
5267
1H NMR (CDCl3) d: 2.21 (t, 2H, J = 7); 2.53–2.65 (m, 2H); 4.16 (s,
2H); 4.40 (t, 2H, J = 7); 6.68 (dd, 1H, J = 2 Hz, J = 4 Hz); 7.02–7.14
(m, 5H); 7.21–7.33 (m, 4H); 7.80–7.98 (m, 4H); 8.00 (d, 1H,
J = 4 Hz); 8.86 (s, 1H); 11.07 (br s, 1H). MW 527.58. Anal. Calcd
for C30H23N7O2: C, 70.57; H, 4.78; N, 18.58. Found: C, 70.30; H,
4.56; N, 18.74.
modifications.50 Membranes were incubated with about
150,000 cpm of [
32P]ATP for 20 min in the incubation mixture
a
-
as described without EGTA and NaCl. For agonists the EC50-values
for the stimulation of adenylyl cyclase were calculated with the
Hill equation. Hill coefficients in all experiments were near unity.
IC50 values for concentration-dependent inhibition of NECA-stimu-
lated adenylyl cyclase caused by antagonists was calculated
accordingly. Dissociation constants (Ki) for antagonist were then
calculated with the Cheng and Prusoff equation.51
3.1.1.47. 5-[(4-Fluorobenzyl)carbonyl]amino-8-(3-phenylpro-
pyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine
(180). Yield 75%, yellow solid; mp 90 °C (EtOAc–light petroleum);
IR (KBr): 3230–2990, 1675, 1615, 1580, 1510 cmꢀ1 1H NMR
;
5. Result and discussion
5.1. General strategy
(CDCl3) d: 2.23 (t, 2H, J = 7); 2.51–2.63 (m, 2H); 3.98 (s, 2H); 4.38
(t, 2H, J = 7); 6.72 (dd, 1H, J = 2, J = 4); 7.01–7.58 (m, 10H); 7.96
(d, 1H, J = 2); 8.86 (s, 1H); 11.03 (br s, 1H). MW 495.51. Anal. Calcd
for C27H22N7O2F: C, 65.45; H, 4.48; N, 19.79. Found: C, 65.09; H,
4.36; N, 19.56.
Support Vector Machine represents a group of supervised learn-
ing techniques, firstly applied in pattern recognition, which find
now diverse applications in classification and regression prob-
lems.13,14 SVM has been originally developed in fact for classifica-
3.1.1.48. 5-[(4-Chlorophenoxymethyl)carbonyl]amino-8-(3-
phenylpropyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-
c] pyrimidine (186). Yield 95%, pale yellow solid; mp 166 °C
(EtOAc–light petroleum); IR (KBr): 3225–2985, 1675, 1612,
tion, then the introduction of a suitable e-insensitive loss function
together with the advantages of the kernel representation, have
enabled its application in the regression analysis. In the last years,
the use of SVM approach and SVR has helped to solve several clas-
sification problems, as for example active and non-active com-
pounds discrimination, and to derive quantitative structure–
activity relationships for the prediction of different chemical and
biological properties.15–19,23 SVR seems to be a promising tool, with
good generalization performance and increased robustness com-
pared with the traditional neural networks.52,53
1578, 1513 cmꢀ1 1H NMR (CDCl3) d: 2.24 (t, 2H, J = 7); 2.48–
;
2.59 (m, 2H); 4.38 (t, 2H, J = 7); 5.12 (s, 2H); 6.61 (dd, 1H,
J = 2, J = 4); 6.97 (d, 2H, J = 9); 7.05–7.45 (m, 8H); 7.80 (d, 1H,
J = 2); 8.47 (s, 1H); 10.32 (br s, 1H). MW 527.96. Anal. Calcd
for C26H22N7O3Cl: C, 63.34; H, 4.33; N, 19.15. Found: C,
63.01; H, 4.30; N, 19.27.
3.1.1.49. 5-[(2,6-Dichlorobenzyl)carbonyl]amino-8-(2-phenyl-
propyl)-2-(2-furyl)-pyrazolo[4,3-e]1,2,4-triazolo[1,5-c]pyrimidine
(187). Yield 86%, pale yellow solid; mp 210 °C (EtOAc–light petro-
Furthermore, over the last few years we have stressed the
concept that both topological and electrostatic complementari-
ties are considered two key concepts in the molecular recogni-
tion processes. We consider this conception extremely crucial
in describing the receptor subtypes selectivity. Gasteiger and col-
laborators investigated the MEP on the molecular surface as a
particularly useful method for rationalizing the interactions be-
tween molecules and molecular recognition processes.38,43,44
The introduction of the autocorrelation vector allows then for
overcoming the MEP information inconvenience to be reliant
on the spatial rotation and translation of the molecule. We have
already demonstrated that the autoMEP vectors can be used as
interesting molecular descriptors in different 3D-QSAR applica-
tions.24–28 In this context, we have also reported that pyrazolo-
leum); IR (KBr): 3225–2993, 1679, 1610, 1575, 1508 cmꢀ1 1H
;
NMR (CDCl3) d: 2.25 (t, 2H, J = 7); 2.46–2.58 (m, 2H); 4.40 (t, 2H,
J = 7); 4.99 (s, 2H); 6.62 (dd, 1H, J = 2, J = 4); 7.08–7.40 (m, 9H);
8.20 (d, 1H, J = 2); 9.22 (s, 1H); 11.53 (br s, 1H). MW 546.41. Anal.
Calcd for C27H21N7O2Cl2: C, 59.35; H, 3.87; N, 17.94. Found: C,
59.61; H, 3.94; N, 17.78.
4. Biology
4.1. Binding at human A1, A2A, and A3 adenosine receptors
All pharmacological methods followed the procedures as de-
scribed earlier.49 In brief, membranes for radioligand binding were
prepared from CHO cells stably transfected with human adenosine
receptor subtypes in a two-step procedure. In a first low-speed step
(1000g) cell fragments and nuclei were removed. The crude mem-
brane fraction was sedimented from the supernatant at 100,000g.
The membrane pellet was resuspended in the buffer used for the
respective binding experiments, frozen in liquid nitrogen and
stored at ꢀ80 °C. For the measurement of adenylyl cyclase activity
only one high speed centrifugation of the homogenate was used.
The resulting crude membrane pellet was resuspended in 50 mM
Tris/HCl, pH 7.4 and immediately used for the cyclase assay.
For radioligand binding at A1 adenosine receptors 1 nM
[3H]CCPA was used, whereas 30 and 10 nM [3H]NECA were used
for A2A and A3 receptors, respectively. Non-specific binding of
[3H]CCPA was determined in the presence of 1 mM theophylline,
triazolo-pyrimidine is a versatile scaffold to cover a large
spectrum of the adenosine receptor selectivity. In particular, pyr-
azolo-triazolo-pyrimidines bearing specific substitutions at the
N5 and N8 positions have been described as highly potent and
selective human A3R antagonists while the position N7 shifts
the selectivity profile to the human A2AR subtype.11,12 However,
even if the N5–N7–N8 substitutability rule seems help to work as
a selectivity thumb rule, it is not absolutely trivial to correctly
assigned the selectivity profile of a new pyrazolo-triazolo-pyrim-
idine antagonists.
In this paper, we present an integrated approach based on the
presence of two distinct Support Vector Machine tools both using
as input matrix the above mentioned autoMEP vectors. The first
tool is a SVM-driven selectivity classificator and the second one a
couple of SVR-driven receptor-affinity predictors. The workflow of
the abovementioned procedure is summarized in Figure 1.
The basic idea beyond our approach is to use the autocorrelated
molecular descriptors encoding for the Molecular Electrostatic Po-
tential (autoMEP) of each pyrazolo-triazolo-pyrimidine antagonist
to optimize the experimental A2AR/A3R subtypes selectivity profile
using a SVM classificator (autoMEP/SVMclass). Once a statistically
acceptable autoMEP/SVMclass model has been derived, the two dif-
ferent A2AR and A3R receptor-affinity predictors can be generated
in the case of [3H]NECA 100 M R-PIA was used.49
l
4.2. Functional assay for A2B antagonists
Due to the lack of a suitable A2B radioligand the potency of
antagonists was determined in adenylyl cyclase experiments. The
procedure was carried out as described previously with minor

5268
L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
using even in this case the autoMEP descriptors as input values.
Following our previously experience in developing valuable tools
for the prediction of receptor subtypes selectivity, the use of the
autoMEP/SVMclass model ahead of the two receptor-affinity pre-
dictors can reduce the number of misattributed compounds that
usually derived from the simple ratio between the two predicted
A2AR and A3R receptor-affinity values. In fact, the receptor binding
predictions are typically accompanied by a standard errors that
compromise the accuracy of their ratio values. In principle, for
any external new pyrazolo-triazolo-pyrimidine analog, after calcu-
lating its autoMEPs, it could be possible to simultaneously access at
both A2AR/A3R subtypes selectivity profile and A2A/A3R receptor-
affinity values.
model has been obtained setting the C parameter value to 150
and the parameter value to 0.01 as shown in Figure 2b. This mod-
c
el has been subjected to an extensive n-fold validation procedure
and the results are collected in Table 2a. The model sensitivity
and specificity, computed from the confusion matrix after leave-
one-out cross-validation procedure, are summarized in Table 2b.
The percentages obtained after repeated 10-fold and fivefold
validation processes confirm the statistical reliability of this model.
Interestingly, it yielded around 90% correct predictions after leave-
one-out cross-validation. The statistical robustness is also con-
firmed by the values of sensitivity and specificity. We consider this
percentage amount as an acceptable threshold to validate the
robustness of this model to appropriately infer about hA2AR versus
hA3R selectivity spectrum for novel, but structurally related, antag-
onists. Therefore, we decided to select this model as SVM-driven
selectivity classificator for the final validation step, described in
the validation paragraph.
5.2. SVM classification model (autoMEP/SVMclass)
To construct our SVM-driven selectivity classificator, we have
selected 104 ‘selective’ pyrazolo-triazolo-pyrimidines derivatives
(1–104), and we have defined them as our training set. The selec-
tivity profile (hA2AR versus hA3R) is summarized in Figure 2a.
The definition of ‘selectivity’ used for our classification ap-
proach is based on a simple binary criteria. In particular, the binary
criteria acts as simple on–off switch: the selectivity index is set to
‘+1’ if it is referred to a selective hA2AR antagonist, and is ‘ꢀ1’ for a
selective hA3R antagonists. Moreover, we have considered as selec-
tivity threshold a difference of at least two orders of magnitude be-
tween the corresponding Ki values.
5.3. SVR regression model (autoMEP/SVR)
A different collection of 104 selective pyrazolo-triazolo-pyrimi-
dines analogs (1–71, 105–137) has been selected as training set of
known hA2A and hA3 adenosine receptor antagonists to derive our
autoMEP/SVR model. In this specific case, the selection of all train-
ing set candidates has not necessarily related to their selectivity
profiles due to the fact that the principal aim of our predictor is
to accurately predict the receptor-binding affinity independently
Information encoded by the autoMEP vectors has been used as
input matrix for our SVM classification protocol. Using a Gaussian
radial basis function kernel, several classification models have
been generated analyzing all possible combination of the C param-
to their spectrum selectivity. Indeed, we have utilized 17 hA2AR
selective antagonists, 54 hA3R selective antagonists and 33 non-
selective antagonists, as summarized in Figure 3.
Also in this step, information encoded by the autoMEP vectors
of all training set antagonists has been used as input matrix for
our SVR prediction protocol. For the generation of both regression
eter (from 10 to 150 with an interval of 10) and the
c parameter
(from 0.001 to 0.015 with an interval of 0.001). The optimized
Figure 1. Flowchart of the in series autoMEP/SVMclass and autoMEP/SVR approach for the selection of new selective and potent human A2A and A3 receptor antagonists.

L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
5269
Figure 2. SVMclass finally selected model (104 molecules). (a) Data distribution indicating the percentage % of hA2AR antagonists ( —red) and the percentage% of hA3R
antagonists ( —blue) in the training set (50 hA2AR and 54 hA3R antagonists); (b) parameter selection in SVMclass model: percentage % of correct prediction in leave-one-out
validation procedure selecting different values of C (c = 0.01) and selecting different values of c (C = 150).
Table 2a
AutoMEP/SVMclass model: the statistical parameters after the cross-validation
procedure on the selected classificator are collected
Partition
CV
No. of runs
% Correct predictions
Mean
Stdev
Min
Max
Training set
Training set
1
1
10
10
1
99.0
93.3
91.4
91.7
78.4
—
—
2.5
2.1
—
99.0
93.3
86.5
88.5
78.4
99.0
93.3
95.2
94.2
78.4
LOO
10-fold
Fivefold
Test set
Figure 3. Data distribution representation indicating the percentage % of selective
hA2AR antagonists ( —red), the percentage % of selective hA3R antagonists (
—
blue) and the percentage % of non-selective compounds ( —green): (a) in the
training set of SVR hA2AR and SVR hA3R models (17 hA2AR and 54 hA3R antagonists,
33 non-selective antagonists); (b) in the test set (10 hA2AR and 5 hA3R antagonists,
36 non-selective antagonists, including 17 more active on hA2AR antagonists and 19
more active on hA3R antagonists).
Table 2b
AutoMEP/SVMclass model: the statistical parameters after the LOO cross-validation
procedure are reported
models are notably better than the results of the respective RSA se-
lected predictors. See Supplementary data for details.
Correct predictions (%)
93.3
Sensitivity (%)
92.0
Specificity (%)
94.4
LOO-cv
5.4. Validation of the tandem autoMEP/SVMclass and autoMEP/
SVR models
models, also in this case we have utilized a Gaussian radial basis
function kernel. In the case of A2A model several prediction models
have been generated analyzing all possible combination of the C
As anticipated, the principal aim of the present work has been
to evaluate the robustness of the tandem autoMEP/SVMclass and
autoMEP/SVR in the prediction of both A2AR/A3R subtypes selectiv-
ity and receptor binding affinity profiles of new pyrazolo-triazolo-
pyrimidines derivatives.
parameter (from 50 to 200 with an interval of 10), the
(from 0.0005 to 0.001 with an interval of 0.0001 and from 0.001 to
0.005 with an interval of 0.001) and the parameter (from 0.0005
e parameter
c
To achieve to this objective, 51 novel pyrazolo-triazolo-pyrimi-
dines analogs (138–188) have been synthetized and pharmacolog-
ically characterized. The experimental A2AR and A3R binding
affinities are collected in Table 3. The human A1R binding affinity
and the IC50 referring to the inhibition of NECA-stimulated aden-
ylyl cyclase activity in CHO cells expressing human A2B receptors
are reported in the Table 1. Moreover, the experimental spectrum
of selectivity is summarized in Figure 3. We also performed a Prin-
cipal Component Analysis (PCA) to explore the similarity between
the compounds used as training set in the autoMEP/SVMclass mod-
el and the novel selective hA2AR (molecules 142, 153, 155, 156,
159, 170, 172, 174–176 in Table 3) and hA3R antagonists (mole-
cules 139, 141, 145, 152, 181 in Table 3) in the test set.54 Chemical
information encoded in both hA2AR and hA3R antagonists was
compared by considering the autoMEP vectors as variables in the
PCA. According to this analysis, the closer two molecules are in
the score plot, the higher is their similarity with respect to the
components concerned. Figure 6 shows the overlap between the
training set and test set chemical spaces.55
to 0.001 with an interval of 0.0001 and from 0.001 to 0.005 with an
interval of 0.001). Slightly different combinations of these param-
eters have been used for the generation of the optimized hA3R
SVR model (the C parameter has been changed from 50 to 200 with
an interval of 10, the
e
parameter from 0.1 to 0.9 with an interval of
0.1 and the
0.01).
c parameter from 0.005 to 0.05 with an interval of
An acceptable hA2AR SVR model (C = 200, e = 0.0005, c = 0.0006,
as showed in Fig. 4) has been obtained as indicated by the leave-
one-out cross-validated correlation coefficient (rcv) of 0.78 and a
root mean square of residuals (RMSR) of 0.050. autoMEP/SVR sta-
tistical parameters are summarized in Figure 4.
On the other hand, the best hA3R SVR predictor (C = 150,
e = 0.3,
c
= 0.005, as reported in Fig. 5) has been obtained as described by
the leave-one-out cross-validated correlation coefficient (rcv) of
0.85 and a root mean square of residuals (RMSR) of 0.046. auto-
MEP/SVR statistical parameters are summarized in Figure 5.
The results of SVR analysis are noteworthy considering that the
same training set has been used to generated two different robust
models. Both predictors in the validation step are described in the
following paragraph. An alternative non-linear modeling strategy,
Response Surface Analysis (RSA), has been performed using the
same training and test sets. Even if the models robustness is simi-
lar, the SVR predictions on the test set for both hA2AR and hA3R
Following the workflow reported in Figure 1, the autoMEP vec-
tors of these new 51 antagonists have been used as input matrix
for the previously generated autoMEP/SVMclass classificator. As
previously shown in Table 2a, our classification model was able
to correctly assign the 78% of the compounds in the collected test

5270
L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
P
Figure 4. AutoMEP/SVR hA2A
R
model results. (a) Cross-validated
r
after leave-one-out procedure: rcv = [SXY/(SXX)1/2(SYY)1/2], SXY = (X ꢀ Xmean)(Y ꢀ Ymean),
P
P
2
2
SXX = (X ꢀ Xmean
)
and SYY = (Y ꢀ Ymean
) with X = Yexperimental and Y = Ypredicted; (b) RMSR: root mean square of residuals.
P
Figure 5. AutoMEP/SVR hA3R model results. (a) Cross-validated
r
after leave-one-out procedure: rcv = [SXY/(SXX)1/2(SYY)1/2], SXY = (X ꢀ Xmean)(Y ꢀ Ymean),
P
P
2
2
SXX = (X ꢀ Xmean
)
and SYY = (Y ꢀ Ymean
) with X = Yexperimental and Y = Ypredicted; (b) RMSR: root mean square of residuals.

Table 3
Predicted and experimental hA2AR and hA3R both pKi and Ki for the test set
Differences between predicted and experimental pKi values for both SVR models (hA2AR and hA3R) are reported.
*
Predicted pKi—experimental pKi.
In the middle column hA2AR selective antagonists ( —red) and hA3R selective antagonists ( —blue) are highlighted.

5272
L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
Table 4b
set to their class. In more detail, a sensitivity of 92.0% and a spec-
ificity of 65.4% were achieved by predicting the test set. Surpris-
ingly, almost all the selective compounds of our test set are
correctly classified and only 11 of them (22%) have been errone-
ously recognized (Tables 4a, and b).
autoMEP/SVMclass model: both predicted and experimental hA3R classification
results are reported for the classified hA3R compounds in the test set
Analyzing in more detail these 11 derivatives, we can consider
these as ‘tricky’ cases considering both the entity of the corre-
sponding experimental errors, and the similarity between their
hA2AR and hA3R Ki values. Even if preliminary, we consider the per-
formance of this selectivity classificator very encouraging.
After passing the selectivity filtering process, each of the hA2A
R
and hA3R antagonist has been analyzed by the corresponding SVR
binding affinity predictor. The comparison of all the experimental
with the predicted pKi values by the abovementioned hA2AR and
hA3R SVR models on the test set again support the quality of the
predictors, as underlined by the good values of the correlation
coefficient (q = 0.82 and q = 0.85, respectively) (Figs. 6a and 7aa).
In Figures 6b and 7b only the hA2A classified antagonists predicted
by the hA2AR SVR model and only the hA3A classified antagonists
predicted by the hA3R SVR model respectively have been reported.
The prediction accuracies, as demonstrated by the differences be-
tween the experimental and the predicted pKi, are statistically sat-
isfying for both hA2AR and hA3R SVR models, with few exceptions
in particular regarding the hA3R binding affinity prediction such as
168, 170, 185 and 188 (see Table 3, Figs. 7b and 8b).
In the first column the correctly classified hA3R antagonists—true positive—(
green) are highlighted.
—
5.5. Structure–activity relationship analysis
All newly synthesized pyrazolo-triazolo-pyrimidines (138–188)
were tested against all four human adenosine receptor subtypes.
The data summarized in Table 1 show that all the tested com-
pounds 138–188 display a broad affinity spectrum against the four
adenosine receptor subtypes with different levels of receptor sub-
types selectivity.
It is clearly evident, as previously observed for the pyrazolo-
triazolo-pyrimidines as human A3 adenosine receptor antagonists,
that the presence of a methyl group at the N8 position seems to
confer a better profile at the hA3 adenosine receptor with an affin-
ity in the low nanomolar range independently from the substitu-
tion at the N5 position (e.g., compounds 147, 150, 152, 161–164,
168) with good levels of selectivity vs the other receptor subtypes.
The effect of the methyl substitution at the N8 position is well de-
scribed by both autoMEP/SVMclass and autoMEP/SVR models, as
highlighted in Tables 3, 4a and b.
Figure 6. Graphical representation of the scores resulting from the Principal
Component Analysis (PCA) by using 12 autoMEP vectors for each molecule as
variables. The compounds included in the training set of the autoMEP/SVMclass
classificator ( —green), selective hA2AR antagonists ( —red) and hA3R antagonists
Interestingly, a quitedifferentbindingprofilewasobserved when
(
—blue) in the test set are plotted. The projections for each compound on the first
at the N8 position bulkier groups than a methyl substituent were
three principal components have been considered.
Table 4a
AutoMEP/SVMclass model: both predicted and experimental hA2AR classification results are reported for the classified hA2AR compounds in the test set
In the first column the correctly classified hA2AR antagonists—true positive—( —green) are highlighted.

L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
5273
Figure 7. (a) Experimental pKi activity data plotted versus Predicted pKi values by the autoMEP/SVR hA2AR model for the test set; (b) experimental pKi activity data (j—black)
of the classified selective hA2AR antagonists included in the test set compared to pKi values predicted by autoMEP/SVR hA2AR model ( —red).
Figure 8. (a) Experimental pKi activity data plotted versus Predicted pKi values by the autoMEP/SVR hA3R model for the test set; (b) experimental pKi activity data (j—black)
of the classified selective hAA3R antagonists included in the test set compared to pKi values predicted by autoMEP/SVR hA3R model ( —blue).
introduced. In fact, when ethyl, propyl orn-butylgroupsare linkedat
theN8 pyrazolenitrogenincombinationwithseveralarylacetylmoi-
eties at the N5 position, a major change in binding affinities were de-
tected and consequently a shift of selectivity spectrum was also
observed. In particular, the presence of a 2-thienyl or 3-thienyl
groups at the N5 position significantly increase the affinity at the
hA2AR while a significant reduction of affinity at the hA3R was
observed, with a resulting increase of selectivity vs this receptor
subtype (e.g., compound 170: KihA1 = 25.5 nM, KihA2A = 3.13 nM,
IC50hA2B = 906 nM, KihA3 = 189 nM). Also in this specific case, our
predictions seem to be very closed to the experimental data.
of compound 141 (KihA1 = 441 nM, KihA2A = 159 nM, IC50hA2B =
>10,000 nM, KihA3 = 5.86 nM), which still retains good affinity for
the hA3 adenosine receptor. A quite similar behavior could be de-
tected when bulky substituents (phenylethyl or phenylpropyl)
are linked at the N8 position: In fact in general, with the exception
of some substituents at the N5 position (e.g., diphenylacetyl: com-
pounds 143 and 181), all the derivatives showed a binding profile
favorable to the A2A adenosine receptors but the levels of selectiv-
ity are very low.
In contrast with our previous studies, it seems that a combina-
tion of bulky substituents at the N5 and N8 positions (e.g., com-
pounds 138 and 154) resulted to be detrimental in term of
affinity at the four adenosine receptors subtypes and in particular,
as theoretically demonstrated, at the A2B adenosine receptors. Also
in this case, a good correspondence between predicted and exper-
imental results has been found.
Another interesting convergent scenario between predicted and
experimental results regards the influence on the affinity of the
naphthylacetyl moiety at the N5 position. In fact comparing the
a
and b isomers it is clearly evident that while the
good potency and selectivity for the hA3 adenosine receptor (e.g.,
compound 157: KihA1 = 596 nM, KihA2A = 74.6 nM, IC50hA2B
a-isomer retains
=
50,000 nM, KihA3 = 3.05 nM), the b-isomer resulted to be more
potent for the hA2A subtype (e.g., compound 158: KihA1 = 185 nM,
KihA2A = 18.4 nM, IC50hA2B = 10,370 nM, KihA3 = 36.3 nM).
Among this series, significant differences in terms of affinity
could be observed when at the N8 position a branched chain, such
as isopentyl, is present. In fact, all the derivatives bearing this sub-
stitution at the N8 position, independently from the substitutions
at the N5 position showed the best affinity versus the hA2A adeno-
6. Conclusions and perspectives
In the present work, we have presented a combination of Sup-
port Vector Machine tools able to predict both A2AR versus A3R
subtypes selectivity profile and the corresponding receptor binding
affinities of a large dataset of known pyrazolo-triazolo-pyrimidine
analogs. The preliminary results based on a new set of 51 pyrazolo-
triazolo-pyrimidines are very encouraging. To further validate our
integrated SVM approach, we are extending the applicability of this
method to other class of A2AR and A3R antagonists and, at the same
sine receptors (e.g., compound 175: KihA1 = 48.2 nM, KihA2A
=
5.43 nM, IC50hA2B = 5480 nM, KihA3 = 266 nM) with the exception

5274
L. Michielan et al. / Bioorg. Med. Chem. 17 (2009) 5259–5274
20. Liu, H. X.; Zhang, R. S.; Yao, X. J.; Liu, M. C.; Hu, Z. D.; Fan, B. T. J. Chem. Inf.
Comput. Sci. 2004, 44, 161.
21. Norinder, U. Neurocomputing 2003, 55, 337.
22. Xue, C. X.; Zhang, R. S.; Liu, H. X.; Yao, X. J.; Liu, M. C.; Hu, Z. D.; Fao, B. T. J. Chem. Inf.
Comput. Sci. 2004, 44, 1693.
time, we are exploring the possibility to achieve, using a multi-
classificator, to the full adenosine selectivity spectrum (from the
human A1R to the A3 subtype).
23. Oloff, S.; Mailman, R. B.; Tropsha, A. J. Med. Chem. 2005, 48, 7322.
24. Moro, S.; Bacilieri, M.; Ferrari, C.; Spalluto, G. Curr. Drug Discovery Technol.
2005, 2, 13.
Acknowledgments
25. Moro, S.; Bacilieri, M.; Cacciari, B.; Spalluto, G. J. Med. Chem. 2005, 48, 5698.
26. Moro, S.; Bacilieri, M.; Cacciari, B.; Bolcato, C.; Cusan, C.; Pastorin, G.; Klotz, K.
N.; Spalluto, G. Bioorg. Med. Chem. 2006, 14, 4923.
27. Bacilieri, M.; Kaseda, C.; Spalluto, G.; Moro, S. Lett. Drug Des. Discovery 2007, 4,
122.
28. Michielan, L.; Bacilieri, M.; Schiesaro, A.; Bolcato, C.; Pastorin, G.; Spalluto, G.;
Cacciari, B.; Klotz, K. N.; Kaseda, C.; Moro, S. J. Comput. Inf. Model. 2008, 48, 350.
29. OPENMOSIX, Version 2.4.26; Moshe Bar, Tel Aviv University, Israel, 2004.
The molecular modeling work coordinated by S.M. was carried
out with financial support from the University of Padova, Italy, and
the Italian Ministry for University and Research (MIUR), Rome,
Italy. We thank the Molecular Network GmbH (Erlangen, Germany)
for the assistance in using the Adriana modeling suite. S.M. is also
very grateful to the Chemical Computing Group for the scientific
and technical partnership.
30. ADRIANACODE; Version 2.0; Molecular Networks GmbH: Erlangen, Germany, 2005.
light
31. Joachims, T. SVM
; Version 6.01; Support Vector Machine. http://
svmlight.joachims.org, 2004.
Supplementary data
32. Baraldi, P. G.; Cacciari, B.; Romagnoli, R.; Spalluto, G.; Moro, S.; Klotz, K. N.;
Leung, E.; Varani, K.; Gessi, S.; Merighi, S.; Borea, P. A. J. Med. Chem. 2000, 43,
4768.
33. Baraldi, P. G.; Cacciari, B.; Romagnoli, R.; Spalluto, G.; Varani, K.; Gessi, S.;
Merighi, S.; Borea, P. A. Drug Dev. Res. 2001, 52, 406.
34. Baraldi, P. G.; Cacciari, B.; Romagnoli, R.; Spalluto, G.; Monopoli, A.; Ongini, E.;
Varani, K.; Borea, P. A. J. Med. Chem. 2002, 45, 115.
35. Baraldi, P. G.; Cacciari, B.; Spalluto, G.; Bergonzoni, M.; Dionisotti, S.; Ongini, E.;
Varani, K.; Borea, P. A. J. Med. Chem. 1998, 41, 2126.
Supplementary data (Training sets (Tables 1 and 2) and test set
(Table 3), together with Table 4 of elemental analyses of synthe-
sized compounds) associated with this article can be found, in
the online version, at doi:10.1016/j.bmc.2009.05.038.
References and notes
36. Baraldi, P. G.; Fruttarolo, F.; Tabrizi, M. A.; Preti, D.; Romagnoli, R.; El-Kashef,
H.; Moorman, A.; Varani, K.; Gessi, S.; Merighi, S.; Borea, P. A. J. Med. Chem.
2003, 46, 1229.
37. Baraldi, P. G.; Cacciari, B.; Romagnoli, R.; Klotz, K. N.; Spalluto, G.; Varani, K.;
Gessi, S.; Merighi, S.; Borea, P. A. Drug Dev. Res. 2001, 53, 225.
38. Gasteiger, J.; Li, X.; Rudolph, C.; Sadowski, J.; Zupan, J. J. Am. Chem. Soc. 1994,
116, 4608.
39. Gasteiger, J.; Marsili, M. Tetrahedron 1980, 36, 3219.
40. Gasteiger, J.; Saller, H. Angew. Chem., Int. Ed. Engl. 1985, 24, 687.
41. Moreau, G.; Broto, P. Nouv. J. Chim. 1980, 4, 757.
1. Moro, S.; Bacilieri, M.; Deflorian, F.; Spalluto, G. New J. Chem. 2006, 30, 301.
2. Moro, S.; Deflorian, F.; Bacilieri, M.; Spalluto, G. Curr. Pharm. Des. 2006, 12,
2175.
3. Fredholm, B. B.; Arslan, G.; Halldner, L.; Kull, B.; Schulte, G.; Wasserman, W.
Naunyn-Schmiedeberg’s Arch. Pharmacol. 2000, 362, 364.
4. Jacobson, K. A.; Gao, Z. G. Nat. Rev. Drug Discovery 2006, 5, 247.
5. Muller, C. E. Curr. Top. Med. Chem. 2003, 3, 445.
6. Jacobson, K. A. In Annual Reports in Medicinal Chemistry; Elsevier: San Diego, CA,
2003.
42. Moreau, G.; Broto, P. Nouv. J. Chim. 1980, 4, 359.
43. Wagener, M.; Sadowski, J.; Gasteiger, J. J. Am. Chem. Soc. 1995, 117, 7769.
44. Bauknecht, H.; Zell, A.; Bayer, H.; Levi, P.; Wagener, M.; Sadowski, J.; Gasteiger,
J. J. Chem. Inf. Comput. Sci. 1996, 36, 1205.
7. Johnston, T. H.; Brotchie, J. M. Curr. Opin. Invest. Drugs 2006, 7, 25.
8. Ferre, S.; von Euler, G.; Johansson, B.; Fredholm, B. B.; Fuxe, K. Proc. Natl. Acad.
Sci. U.S.A. 1991, 88, 7238.
45. Vapnik, V. The Nature of Statistical Learning Theory; Springer: New York, 1995.
46. Vapnik, V. Statistical Learning Theory; Wiley: New York, 1998.
47. Smola, A. J.; Scholkopft, B. Learning with Kernels. Support Vector Machines,
Regularization, Optimization, and Beyond; MIT Press, 2002.
48. Smola, A. J.; Scholkopft, B. A Tutorial on Support Vector Regression; NeuroCOL
Technical Report Series, NCR2-TR-1998-030, 1998.
9. Ribeiro, J. A.; Sebastiao, A. M.; de Mendonca, A. Prog. Neurobiol. 2002, 68, 377.
10. Xu, K.; Bastia, E.; Schwarzschild, M. Pharmacol. Ther. 2005, 105, 267.
11. Baraldi, P. G.; Cacciari, B.; Moro, S.; Spalluto, G.; Pastorin, G.; Da Ros, T.; Klotz,
K. N.; Varani, K.; Gessi, S.; Borea, P. A. J. Med. Chem. 2002, 45, 770.
12. Baraldi, P. G.; Tabrizi, M. A.; Bovero, A.; Avitabile, B.; Preti, D.; Fruttarolo, F.;
Romagnoli, R.; Varani, K.; Borea, P. A. Eur. J. Med. Chem. 2003, 38, 367.
13. Burges, C. J. C. Data Min. Knowl. Discovery 1998, 2, 121.
14. Cristianini, N.; Shawe-Taylor, J. An Introduction to Support Vector Machines;
Cambridge University Press, 2000.
15. Czerminski, R.; Yasri, A.; Hartsough, D. Quant. Struct. Act. Relat. 2001, 20, 227.
16. Warmuth, M. K.; Liao, J.; Ratsch, G.; Mathieson, M.; Putta, S.; Lemmen, C. J.
Chem. Inf. Comput. Sci. 2003, 43, 667.
17. Jorissen, R. N.; Gilson, M. K. J. Chem. Inf. Model. 2005, 45, 549.
18. Bruce, C. L.; Melville, J. L.; Pickett, S. D.; Hirst, J. D. J. Chem. Inf. Model. 2007, 47,
219.
49. Li, A. H.; Moro, S.; Forsyth, N.; Melman, N.; Ji, X. D.; Jacobson, K. A. J. Med. Chem.
1999, 42, 706.
50. Klotz, K. N.; Hessling, J.; Hegler, J.; Owman, C.; Kull, B.; Fredholm, B. B.; Lohse,
M. J. Naunyn-Schmiedeberg’s Arch. Pharmacol. 1998, 357, 1.
51. Cheng, Y. C.; Prusoff, H. R. Biochem. Pharmacol. 1973, 22, 3099.
52. Demiriz, A.; Bennett, K. P.; Breneman, C. M.; Embrechts, M. J. Support Vector
Regression in Chemoinformatics In Computing Science and Statistics: Proceeding of
Interface; 2001.
53. Ivanciuc, O. Rev. Comput. Chem. 2007, 23, 291.
54. THE UNSCRUMBLER, Version 9.8; CAMO Process AS: Oslo, Norway, 2008.
55. MOLEGRO DATA MODELLER, Version 1.0.0; Molegro ApS: Aarhus C, Denmark, 2007.
19. Liu, H. X.; Zhang, R. S.; Yao, X. J.; Liu, M. C.; Hu, Z. D.; Fan, B. T. J. Chem. Inf.
Comput. Sci. 2003, 43, 1288.