Improving the inhibitory activity of arylidenaminoguanidine compounds at the N-methyl-d-aspartate receptor complex from a recursive computational- experimental structure-activity relationship study

Article abstract of DOI:10.1016/j.bmc.2013.01.051

Using a combination of both the partial least squares (PLS) and back-propagation artificial neural network (ANN) pattern recognition methods, several models have been developed to predict the activity of a series of arylidenaminoguanidine analogs as inhib

Full text of DOI:10.1016/j.bmc.2013.01.051

Bioorganic & Medicinal Chemistry 21 (2013) 1764–1774
Contents lists available at SciVerse ScienceDirect
Bioorganic & Medicinal Chemistry
journal homepage: www.elsevier.com/locate/bmc
Improving the inhibitory activity of arylidenaminoguanidine
compounds at the N-methyl-D-aspartate receptor complex
from a recursive computational-experimental structure–activity
relationship study
Joshua R. Ring a, , Fang Zheng a, , Aaron J. Haubner a, John M. Littleton a, Peter A. Crooks a,b,
⇑
a Department of Pharmaceutical Sciences, College of Pharmacy, University of Kentucky, Lexington, KY 40536, USA
b Department of Pharmaceutical Sciences, College of Pharmacy, University of Arkansas for Medical Sciences, Little Rock, AR 72205, USA
a r t i c l e i n f o
a b s t r a c t
Article history:
Using a combination of both the partial least squares (PLS) and back-propagation artificial neural network
(ANN) pattern recognition methods, several models have been developed to predict the activity of a series
of arylidenaminoguanidine analogs as inhibitory modulators of the N-methyl-D-aspartate receptor com-
Received 14 November 2012
Revised 16 January 2013
Accepted 23 January 2013
Available online 31 January 2013
plex. This was done by correlating structural and physicochemical descriptors obtained from computa-
tion software with the experimentally observed [3H]MK-801 displacement ability of a small library of
synthesized and in vitro screened arylidenaminoguanidines. Results for the generated PLS model were
r2 = 0.814, rmsd = 0.208, rC2V = 0.714, loormsd = 0.261. The ANN model was created utilizing the eleven
descriptors from the PLS model for comparison. The quality of the ANN model (r2=0.828, rmsd = 0.200,
r2CV = 0.721, loormsd = 0.257) is similar to the PLS model, and indicates that the feature between the
inputs and the output is majorly linear. These computational models were able to predict inhibition of
the NMDA receptor complex by this series of compounds in silico, affording a predictive structure-based
‘pre-screening’ paradigm for the arylideneaminoguanidine analogs.
Keywords:
Alcohol abuse
N-Methyl-D-aspartate receptor
Arylidenaminoguanidine analogs
Computational modeling
QSAR
Ó 2013 Elsevier Ltd. All rights reserved.
1. Introduction
Regulatory sites on the NMDA receptor are abundant, and the
list of modulators binding at sites outside the receptor includes
The N-methyl-D-aspartate (NMDA) receptor complex has be-
L-glutamate and glycine, the endogenous co-agonists,
D-serine
come one of the most important and widely studied targets for
drug development due to its importance in the central nervous sys-
tem, as well as its complexity and large number of modulatory
molecules and their bindings sites, and hence individual drug tar-
gets. NMDA receptors are the only glutamate receptors that are
highly permeable to calcium upon opening, and they govern
important physiological functions, such as synapse formation, neu-
ronal differentiation, and long-term potentiation, as well as being
implicated in memory and special learning function. However,
the NMDA cation channels are also susceptible to over-activation,
causing excess intracellular Ca2+ concentrations, initiating a neuro-
degenerative cascade leading to cell death1–3 and a number of neu-
rological diseases such as epilepsy4,5 as well as Alzheimer’s6,7 and
Huntington diseases.8 Additionally, antagonism of the NMDA
channel has also been shown to reduce neurotoxicity associated
with alcohol dependence and withdrawal.9–11
and N-methyl- -aspartate, agonists, ifenprodil, an antagonist, and
D
a variety of polyamines, which have varying effects from agonist
to antagonist activity. Many potent NMDA channel antagonists,
such as dextromethorphan, MK-801, phencyclidine (PCP), and ket-
amine, have very high abuse potential and exhibit neurotoxicity.
Polyamines exhibit their NMDA receptor modulatory activity
based on several key structural features, including number of ami-
no-groups, distance between the amino-groups, lipophilic or
hydrophilic domains, and charge. Both spermine and spermidine
act as channel agonists, and potentiate binding of open-channel
blockers. Antagonists at the polyamine sites include diethylenetri-
amine, which serves to block or displace spermidine and sperm-
ine12
, and arcaine (1,4-diguanidinobutane), which displaces
MK-801 from the NMDA channel in the absence or presence of
spermidine.13,14 Agmatine (4-amino-1-guanidino- butane), a poly-
amine synthesized in the mammalian brain15–19, has been shown
to be a competitive antagonist at one or more polyamine sites on
the NMDA receptor complex, therefore being relatively inactive
under physiological conditions and concentrations.20 Studies have
shown that agmatine, although a very polar molecule crosses the
⇑
Corresponding author. Tel.: +1 501 686 6495; fax: +1 501 526 5945.
E-mail address: pacrooks@uams.edu (P.A. Crooks).
These authors contributed equally to this work.
0968-0896/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.bmc.2013.01.051

J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
1765
blood–brain barrier.21 An endogenous ligand at imidazoline bind-
ing sites22, this compound has also been shown to have effects
on learning and memory in rodents.23 The intriguing properties
of agmatine as a potent and selective allosteric competitive inhib-
itor of the NMDA receptor complex have led to interest in optimiz-
ing this molecule for use in drug development.
This present report describes a recursive computational-exper-
imental structure–activity relationship study that may have utility
in improving the inhibitory activity of a series of novel arylidenea-
moidal dose-response curve. All data analysis was performed using
GraphPad PrismÒ (GraphPad Software Inc., San Diego CA).
2.3. Generation of the molecular data sets
Molecular structures were modeled in silico with the use of the
Sybyl discovery software 7.025. Each molecule was constructed
using the program, and formal charges were assigned. The geome-
try optimization (energy minimization) was first performed by
using the molecular mechanics (MM) method with the Tripos force
field and the default convergence criterion. To consider various
possible conformation of a molecule, a variety of conformations
for a given molecule were obtained that were associated with local
minima on the potential energy surface from RANDOMSEARCH in
Sybyl. Each of these conformations was geometrically optimized
with semi-empirical molecular orbital (MO) energy calculations
at the PM3 level to locate the most stable conformation, that is,
the one with the lowest energy.25
An initial set of 27 molecules and a second set of 45 molecules
constituted the dataset for the structure–activity relationship
analysis in this study. All compounds in this study were synthe-
sized in-house, and the experimental IC50 values were measured
in our own laboratory. Initially, 27 molecules were synthesized
and the dataset from these compounds (T1-T27) was used for
model training and leave-one-out (LOO) validation. The trained
model was then used to predict a dataset of 18 additional mole-
cules (P1-P18) and the results analyzed. These 18 new molecules
were then incorporated into the original set of 27 molecules to
form a new set of 45 molecules (Table 2). New QSAR models were
built based on the larger dataset. Experimental IC50 values for
each of the above 72 compounds were obtained providing the
pharmacological parameter which characterizes the interaction
with the NMDA receptor complex. For the set of 27 molecules uti-
minoguanidine compounds at the N-methyl-D-aspartate receptor
complex.
2. Methods
2.1. Synthesis of arylideneaminoguanidine analogs
Novel arylideneaminoguanidine analogs (Scheme 1) were pre-
pared from the reaction of an appropriate aryl aldehyde (1 equiv)
with aminoguanidine hydrochloride (1.25 equiv). These reactants
were dissolved in a minimal amount of methanol, heated to reflux,
and stirred for a period of 2–3 h. Upon cooling to room tempera-
ture, the product crystallized from the methanol solution. Products
were then placed under vacuum overnight to remove any remain-
ing solvent. Experimental details and structural characterization of
the reaction products are described elsewhere.24
2.2. [3H]MK-801 binding assay
Assays were performed at room temperature in 96-well micro-
titer plates with a total capacity of 350
was added 100 L P2 membrane, 100
[3H]MK-801 ligand (6 nM) for a total of 300
l
l
L per well. To each well
L sample and 100
L, with a 15-min
l
lL
l
incubation period between the additions of the sample and ligand.
Three replicates of each of the following were performed per plate:
lized, the IC50 values of 10 of the molecules were 6 50
ecules had IC50 values in the range 50–100 M, and 8 molecules
had IC50 values P100 M; for the set of the 45 molecules, the
IC50 values of 24 molecules were 6 50 M, 12 molecules had
IC50 values in the range 50-100 M, and 9 molecules had IC50
values P100 M.
lM, 9 mol-
l
a control with 100
tromethorphan (3 mM) to measure non-specific binding, and
100 L spermidine (300 M) as a spermidine-potentiated control.
For the preliminary screening, four replicates of each sample
were performed at concentrations of 6 mM and 600 M, which
yielded total well concentrations of 1.0 mM and 100 M. Further
lL Tris acetate buffer as the sample, 100 lL dex-
l
l
l
l
l
l
l
l
2.4. Generation of molecular descriptors
testing to obtain dose-response curves was performed with three
replicates of seven concentrations spanning an ample region to cal-
culate accurate IC50-values from the full dose-response curves,
For all compounds, the optimum three-dimensional conforma-
tions were used for generation of descriptors, some of which were
geometry-dependent. A total of 313 molecular descriptors, consist-
ing of zero-dimensional (constitutional descriptors), one-dimen-
sional (functional groups, empirical descriptors, physical
properties), two-dimensional (topological descriptors), as well as
three-dimensional (geometrical, WHIM, GETAWAY, aromaticity
using 100
lL spermidine (300
lM) in one set for a potentiated
curve and 100
lL buffer in another set to obtain a non-potentiated
curve. All micro-titer plates were allowed to incubate for 40 min
following [3H]MK-801 addition, and were then harvested into filter
plates, which were then dried overnight in the hood at ambient
temperature. 35 lL of Microscint 20 was added to each membrane
on the filter plate; each plate was sealed and then incubated for 2 h
descriptors) variables, were created by the DRAGON program.26
A
reduced set of 134 descriptors was obtained after the constant
and near constant descriptors and the highly inter-correlated
(>0.95) descriptors were discarded.
before reading on a plate counter.
The compounds previously described were evaluated for their
ability to displace [3H]MK-801 in the presence and absence of sper-
midine. For each compound, the two curves were then plotted, and
IC50-values were derived. For binding data, IC50 values were deter-
mined using non-linear regression to a one-site variable-slope, sig-
2.5. Partial least squares regression analysis
Partial least squares (PLS) analysis27 was performed using the
QSAR module of Sybyl version 7.0,25 with the NIPALS algorithm
to extract the original variable into PLS components. All variables
were initially auto-scaled to zero mean and unit variance. This
method produces new variables by a linear combination of the ori-
ginal descriptors and uses them to predict the biological activities.
The advantage of this method is that the method can be used for
strongly correlated, noisy data with numerous independent vari-
ables (e.g., in modeling data sets where the number of descriptors
greatly exceeds the number of observations).27,28
Cl
NH2
Cl
H2N
NH2
NH2
N
H
NH2
O
H
Ar
N
N
H
Ar
MeOH
H
Scheme 1. Synthetic scheme for formation of arylideneamino-guanidino
compounds

1766
J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
Table 1
Definitions of descriptors used in the 12-descriptor PLS model for the preliminary set of 27 arylidenaminoguanidine compounds
Desc. #
Name
Definition
7
31
35
56
61
C-027
nNHR
GGI1
HATSm
HATS4v
R2u
R—CH—X, instances of a heteroatom (N, S) being covalently bound to a mono-protonated carbon molecule
Number of secondary amines (aliphatic)
Topological charge index of order 1
Leverage-weighted total index/weighted by atomic masses
Leverage-weighted autocorrelation of lag 4/weighted by atomic van der Waals volumes
R autocorrelation of lag 2/unweighted
66
85
86
119
125
127
135
R4v+
R5v+
T(N..N)
G2 m
E2v
R maximal autocorrelation of lag 4/weighted by atomic van der Waals volumes
R maximal autocorrelation of lag 5/weighted by atomic van der Waals volumes
Sum of topological distances between N. . .N
2nd component symmetry directional WHIM index/weighted by atomic masses
2nd component accessibility directional WHIM index/weighted by atomic van der Waals volumes
Polar surface area (includes all O, N, S atoms and covalently bonded H’s) or fragment-based polar surface area
PSA
Table 2
Experimental, fitted and validated pIC50 values for the 27 molecules in the generated PLS model as well as the predicted pIC50 values for 18 new molecules
ꢀ1
ꢀ1
ꢀ1
ꢀ1
(M )
Name
pIC50,OBS
pIC50,fitted
pIC50,loo
Name
pIC50,pred
(M
)
(M
)
(M
)
AG-2
AG-3
AG-4
AG-5
AG-6
AG-8
AG-9
4.02
4.24
4.57
3.74
4.29
4.42
4.99
4.54
4.77
3.67
4.23
4.05
4.15
4.04
4.37
4.64
4.17
4.36
4.11
3.81
4.36
3.65
4.50
3.27
3.31
3.96
3.65
3.89
4.14
4.59
3.62
4.23
4.49
5.00
4.56
4.72
3.80
4.30
3.98
4.21
4.18
4.37
4.67
4.31
4.34
4.07
3.88
4.27
3.89
4.08
3.31
3.48
3.76
3.48
3.71
4.02
4.59
3.48
4.19
4.55
5.11
4.56
4.66
3.96
4.42
3.89
4.33
4.24
4.38
5.06
4.34
4.31
4.03
3.93
4.20
3.99
3.98
3.36
3.62
3.70
4.00
JR-202
JR-203
JR-204
JR-207
JR-209
JR-211
JR-212
JR-213
JR-214
JR-215
JR-217
JR-218
JR-219
JR-220
JR-223
JR-226
JR-227
JR-228
4.29
4.47
4.44
4.04
4.58
4.19
4.55
4.05
4.17
3.67
4.61
4.88
4.07
4.13
4.23
4.15
3.96
4.20
AG-10
AG-12
AG-13
AG-14
AG-15
AG-17
AG-18
AG-19
AG-20
AG-23
AG-24
JR-121
JR-122
JR-124
JR-125
JR-132
JR-146
JR-147
JR-160
JR-197
2.6. Target properties
The input vector was the set of descriptors for each molecule in
the series, as generated by the previous steps. All descriptors and
targets were normalized to the [0,1] interval using the following
formula:
The log(1/IC50) values (with IC50 value in lM) of the set of aryli-
denamino-guanidine compounds were used as the target pharma-
cological criterion to derive the QSARs.
Xij ꢀ Xj
;min
Xi0
¼
ð1Þ
j
Xj ꢀ Xj
;max
;min
where Xij and X0ij represents the original value and the normalized
value of the j-th (j = 1,. . .k) descriptor for compound i (i = 1,. . .n).
Xmin and Xmax represent the minimum and maximum values for
the j-th descriptor. The network was configured with one or more
hidden layers. During the ANN learning process, each compound
in the training set was iteratively presented to the network. That
is, the input vector of the chosen descriptors in normalized form
for each compound was fed to the input units, and the network’s
output was compared with the experimental ‘target’ value. During
one ‘epoch’, all compounds in the training set were presented,
and weights in the network were then adjusted on the basis of
the discrepancy between network outputs and observed log(1/Ki)
values by back-propagation using the generalized delta rule.
2.7. ANN QSAR modeling
Feed-forward, back-propagation-of-error networks were devel-
oped using a neural network C program.27,29–31 Network weights
(Wji(s)) for a neuron ‘j’ receiving output from neuron ‘i’ in the layer
‘s’ were initially assigned random values between ꢀ0.5 and +0.5.
The sigmoidal function was chosen as the transfer function that
generates the output of a neuron from the weighted sum of inputs
from the preceding layer of units. Consecutive layers were fully
interconnected; there were no connections within a layer or be-
tween the input and the output. A bias unit with a constant activa-
tion of unity was connected to each unit in the hidden and output
layers.

J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
1767
2.8. Cross-validation and testing
The 27 molecule models were cross-validated using the ‘leave-
one-out’ approach.27 The generated models were then used to pre-
dict the IC50 values of 18 new molecules. The new molecules were
synthesized and experimental IC50 values measured to test the
model. The 45-molecule models were cross-validated using the
‘leave-one-out’ approach.
2.9. Evaluation of the QSAR models
QSAR models were assessed by Pearson correlation coefficient
r2, root mean square deviation (rmsd), predictive rC2V, and leave-
one-out root mean square deviation (loormsd). r2CV is defined as:
Figure 1. Comparison of observed pIC50 values versus fitted and LOO-validated
pIC50 values using the 12-descriptor PLS model for the initial set of 27 compounds.
SD ꢀ PRESS
r2cv
¼
ð2Þ
SD
where SD is the sum of squared deviations of each measured log(1/
Ki) value from its mean, and PRESS is the sum of squared differences
between actual and predicted values.
3. Results and discussion
3.1. PLS analysis for a set of synthesized 27 compounds and
prediction for new molecules
The modeling work described in this study was performed in
order to lend insight into the structure-activity relationships be-
tween the molecules synthesized and their inhibitory activity
demonstrated by the displacement of [3H]MK-801 from the NMDA
receptor complex. In addition, this study should provide guidance
in the development of more active arylidenaminoguanidine inhib-
itors at this site, while reducing the cost and time associated with
chemical synthesis by indicating which structures should be syn-
thesized based solely on predicted structures and activities.
The group of 27 arylidenaminoguanidino compounds was the
first to be synthesized. These compounds were structurally simi-
lar, and hypothesized to act at the same binding site on the
NMDA receptor complex. These molecules maintained the aro-
matic ‘arylidenaminoguanidine’ moiety in their structure, and
possessed a variety of aromatic moieties as well as different aro-
matic ring-substituents. The pIC50 values of these compounds
spanned the range of ꢀ3.0 to ꢀ5.0, ensuring that there would
be a sufficient difference in activity that might be assignable to
structural and electronic differences. Since this group of com-
Figure 2. Observed versus calculated pIC50 correlation from the 12-discriptor PLS
model with original 27 data points (filled squares) and 18 predicted values for new
molecules (empty squares).
The structures of 18 virtual arylidenaminoguanidine com-
pounds (JR-202 to JR-228) shown in the 5th column of Table 2
were constructed and input into the Sybyl software program, and
the same twelve molecular descriptors defined in Table 1 were
then generated with the DRAGON program. These compounds
were predicted to have activities of the order of 4–5 pIC50, or IC50
values of 10–100 lM. When subsequently synthesized and
screened in the [3H]MK-801 displacement assay the observed
activities of the new set of compounds were compared to the pre-
dicted values from the linear PLS model.
pounds represented
squares method was chosen to model that structure–activity rela-
tionship of the set of compounds.
The regression equation generated, which linearly correlates
the predicted pIC50 to the descriptors, is:
a small molecular set, the partial least
The rmsd between the predicted pIC50 values and the observed
pIC50 values of the newly synthesized set of 18 molecules is 0.625.
As shown in Figure 2, the observed IC50 values of 12 new molecules
matched with their computationally predicted IC50 values within a
reasonable range. Five of the observed IC50 values among 6 new
molecules that differed the most from the predicted values were
compounds which were more potent than predicted. One of the
reasons for this could be because there were a few molecules with
pIC50 values in 5.0–5.5 range in the 27 molecule dataset for build-
ing the PLS model. In addition, while there were six points that dif-
fered in predicted and observed IC50 values by a full order of
magnitude, three of these were the compounds which generated
biphasic dose-response curves.
pIC50 = 4.779 ꢀ (0.212 ꢁ Desc7) ꢀ (0.041 ꢁ Desc31) ꢀ (0.087 ꢁ
Desc35) + (0.051 ꢁ Desc56) + (1.709 ꢁ Desc61) + (0.547 ꢁ Desc66)
+ (11.329 ꢁ Desc85) + (16.609 ꢁ Desc86) ꢀ (0.007 ꢁ Desc119) ꢀ
(8.381 ꢁ Desc125) + (0.656 ꢁ Desc127) ꢀ (0.003 ꢁ Desc135) r2 =
0.898; rmsd = 0.132; rcv2 = 0.815; loormsd = 0.227; F = 30.693
The PLS model consists of six components. The definitions of the
descriptors are shown in Table 1. The fitted pIC50 values and LOO-
validated pIC50 values from this PLS model, and the observed IC50
values from [3H]MK-801 displacement assays are listed in Table 2.
A graphical representation of the observed versus PLS model-
fitted and LOO validated pIC50 values is shown in Figure 1. The
points are well-clustered along the ideal line, with the exception
of one compound, JR-132, which was predicted to possess approx-
imately three times greater inhibitory activity than demonstrated
in the [3H]MK-801 displacement assay.
3.2. PLS analysis for a set of 45 compounds
After adding the 18 newly synthesized and evaluated molecules
into the 27 molecules dataset for the first generation of PLS mod-
eling, PLS and neural network models were then trained based
on the resulting 45-molecule dataset. The second generation

1768
J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
Table 3
Definitions of descriptors used in the 11-descriptor PLS model for the final set of 45 arylidenaminoguanidine compounds
Desc. #
Name
Definition
18
21
33
G3 m
L2v
Gu
3rd component symmetry directional WHIM index/weighted by atomic masses
2nd component size directional WHIM index/weighted by atomic van der Waals volumes
G total symmetry index/unweighted
42
78
93
109
139
167
172
181
nCIC
C-027
MLOGP
H3u
HATS5v
R2v+
R6e
Total number of rings
R—CH—X, instances of a heteroatom (N, S) being covalently bound to a mono-protonated carbon molecule
Moriguchi octanol-water partition coefficient (log P)
H autocorrelation of lag 3/unweighted
leverage-weighted autocorrelation of lag 5/weighted by atomic van der Waals volumes
R maximal autocorrelation of lag 2/weighted by atomic van der Waals volumes
R autocorrelation of lag 6/weighted by atomic Sanderson electronegativities
R maximal autocorrelation of lag 5/weighted by atomic polarizabilities
R5p+
models should better take into account the improved values gener-
ated by the compounds that exhibited a biphasic binding profile
(and also possibly explain the differences seen in comparing the
predicted and observed inhibition of the arylidenaminoguanidine
compounds).
This second PLS model, with eleven descriptors and based on
data and structures from 45 molecules, had the following regres-
sion equation:
pIC50 = 3.052 ꢀ (1.125 ꢁ Desc18) ꢀ (0.180 ꢁ Desc21) + (2.609
ꢁ Desc33) + (0.495 ꢁ Desc42) ꢀ (0.401 ꢁ Desc78) + (0.359 ꢁ
Desc93) ꢀ (0.710 ꢁ Desc109) ꢀ (9.152 ꢁ Desc139) + (19.287 ꢁ
Desc167) + (2.439 ꢁ Desc172) + (18.708 ꢁ Desc181)r2 = 0.814;
rmsd = 0.208; rcv2 = 0.714; loormsd = 0.261; F = 27.772
This is a PLS model with six components. The definitions of the
Figure 3. Comparison of observed pIC50 values versus fitted and LOO-validated
pIC50 values using the 11-descriptor PLS model for the final set of 45 compounds.
descriptors are shown in Table 3. While only one of these descrip-
Table 4
Observed, fitted and LOO-validated pIC50-values of arylidenamino-guanidines from the generated 11-descriptor PLS and neural network models using the set of 45 molecules
Comp. No.
Comp. Code
Structure
Obsd pIC50
PLS-fitted pIC50
PLS-LOO pIC50
NN-fitted pIC50
NN-LOO pIC50
NH2
Cl
NH2
N
1
AG-17
4.14
4.46
4.50
4.47
4.53
N
H
H
H3C
NH2
Cl
NH2
N
2
3
4
5
6
AG-6
4.09
5.22
4.37
4.64
4.36
4.05
4.98
4.26
4.97
4.18
4.07
4.92
4.09
5.08
4.14
4.00
4.98
4.29
4.94
4.22
3.94
4.93
4.10
5.01
4.17
N
H
H
NH2
Cl
NH2
JR-202
JR-212
JR-211
JR-124
N
N
H
H
NH2
Cl
NH2
N
N
H
H
NH2
Cl
NH2
N
N
N
H
H
NH
2
C l
NH
NC
N
H
2
H

J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
1769
Table 4 (continued)
Comp. No.
Comp. Code
Structure
Obsd pIC50
3.64
PLS-fitted pIC50
4.21
PLS-LOO pIC50
4.27
NN-fitted pIC50
4.19
NN-LOO pIC50
NC
NH2
Cl
N
7
JR-125
JR-226
4.27
4.25
N
H
NH2
H
NH2
Cl
NH2
N
8
9
4.14
4.15
3.95
4.26
3.80
4.30
4.01
4.21
MeOOC
N
H
H
MeOOC
NH2
C l
NH2
N
JR-227
4.55
N
H
H
NH2
Cl
NH2
N
10
11
12
AG-20
AG-13
AG-12
4.60
3.78
4.89
4.56
3.84
4.59
4.54
3.85
4.52
4.60
3.77
4.63
4.60
3.77
4.54
N
H
CF3
HO
H
OH
NH2
Cl
N
N
H
NH2
H
NH2
C l
NH2
N
N
H
OMe
MeO
H
NH
2
Cl
N
13
14
15
16
AG-8
4.46
4.54
3.70
3.98
4.31
4.73
3.87
3.85
4.29
4.79
3.90
3.83
4.31
4.69
3.88
3.78
4.28
4.77
3.92
3.76
N
NH
2
H
H
MeO
NH2
Cl
N
AG-10
JR-197
JR-160
N
NH
2
H
OM e
H
NH
2
Cl
NH
N
H N
2
N
H
2
H
H2N
NH2
Cl
N
N
NH2
H
H
CH3
N
H3C
O
NH2
Cl
NH2
17
18
JR-219
JR-122
5.10
3.83
4.65
3.98
4.60
4.00
4.68
3.92
4.90
3.94
N
N
H
H
O
N
NH2
Cl
N
N
H
NH2
H
O
N
O
NH2
Cl
NH2
19
20
JR-121
JR-228
4.11
4.20
4.02
4.48
3.99
4.51
4.01
4.49
3.98
4.01
N
N
H
H
OH
O
H3C
NH2
Cl
NH2
N
N
H
H
(continued on next page)

1770
J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
Table 4 (continued)
Comp. No.
Comp. Code
Structure
Obsd pIC50
PLS-fitted pIC50
PLS-LOO pIC50
NN-fitted pIC50
NN-LOO pIC50
O
O
N
N
O
H3C
NH2
21
JR-223
5.11
4.92
4.81
4.92
4.83
Cl
NH2
N
N
H
H
Cl
NH2
Cl
22
23
24
25
26
27
28
JR-218
JR-217
JR-220
AG-9
5.15
4.74
5.44
4.99
4.13
4.18
4.26
5.09
4.93
4.92
5.32
3.99
4.23
4.42
5.06
4.97
4.85
5.45
3.92
4.24
4.47
5.10
5.00
4.97
5.19
4.00
4.21
4.45
5.08
5.11
4.83
5.24
3.90
4.22
4.57
N
H
NH2
H
NH2
Cl
NH2
N
Cl
N
H
H
Cl
NH2
Cl
N
N
NH2
H
H
Cl
Cl
NH2
Cl
N
N
H
NH2
H
F
NH2
Cl
N
AG-23
AG-18
AG-24
N
NH2
H
H
F
F
NH2
Cl
NH2
N
N
N
H
H
F
NH2
Cl
NH2
N
H
H
N
NH2
Cl
NH2
N
H
29
JR-209
5.07
5.03
4.99
4.96
4.88
H
H
NH2
Cl
NH2
30
31
32
33
AG-4
4.67
5.06
4.58
4.94
4.77
4.81
4.66
5.02
4.81
4.74
4.68
5.05
4.85
4.87
4.61
4.98
4.89
4.70
4.65
5.00
N
N
H
NH2
Cl
NH2
JR-213
JR-203
JR-204
N
N
H
H
N
NH2
Cl
N
H
NH2
CH3
H
H3C
NH2
Cl
NH2
N
N
H
H

J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
1771
Table 4 (continued)
Comp. No.
Comp. Code
Structure
Obsd pIC50
4.41
PLS-fitted pIC50
4.34
PLS-LOO pIC50
4.32
NN-fitted pIC50
4.42
NN-LOO pIC50
O
O
NH2
Cl
NH2
34
AG-19
4.42
N
N
H
H
NH2
O
N
O
Cl
NH2
N
35
36
37
38
AG-15
AG-2
3.96
4.23
3.17
3.71
3.98
4.20
3.32
3.50
3.97
4.13
3.42
3.39
3.93
4.12
3.44
3.53
3.92
4.21
3.52
3.48
O
N
H
H
H
N
NH2
Cl
NH2
N
N
H
H
NH2
Cl
NH2
N
N
N
JR-147
JR-146
N
H
H
N
NH2
Cl
NH2
N
H
H
H
N
NH2
Cl
39
40
41
AG-3
4.43
4.30
4.31
4.29
4.32
4.50
4.25
4.33
4.54
4.27
4.33
4.52
4.01
4.34
4.58
N
N
NH2
H
H
HN
NH2
Cl
JR-207
JR-214
N
N
N
H
NH2
H
CH3
N
NH2
Cl
N
H
NH2
H
H3C
H
N
NH2
Cl
42
JR-215
4.48
4.49
4.50
4.47
4.46
N
N
H
NH2
H
S
NH2
Cl
N
43
44
AG-5
3.93
4.23
3.97
4.21
4.01
4.20
3.92
4.22
3.87
4.22
N
NH2
H
H
NH2
Cl
N
N
AG-14
N
NH2
H
H
H
H
N
NH2
Cl
NH2
Cl
45
JR-132
4.49
4.52
4.67
4.51
4.39
N
NH2
H2N
N
H
H
tors is also found in the earlier PLS model (C-027), many of the
descriptors are very similar, and reflect molecular traits related
to those which were judged to be important in the first set of
descriptors.
molecules with more diversity features), this new set of points
is well-clustered around the ideal diagonal midline, with very
few data points differing by more than half an order of magni-
tude. Specifically, the fitted values for the arylidenaminoguani-
dine compounds are much closer to the observed values, and
the three biphasic compounds are also much better predicted
than in the previous model. The fitted and LOO-validated data
for this model are disclosed in Table 4 along with the experimen-
tally measured values.
For the second partial least squares model, the observed pIC50
-
values generated from the [3H]MK-801 displacement assay were
plotted versus the PLS model-fitted pIC50-values (see Fig. 3).
While the R2 value of this model is lower than that obtained in
the first PLS model (mainly due to an increase in the number of

1772
J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
3.3. ANN analysis for a set of 45 compounds
This second PLS model from the full set of 45 arylidenamino-
guanidine compounds should have higher predictive power than
the first PLS model from the set of 27 compounds as more molec-
ular and affinity information were included. Using the set of eleven
descriptors used in the second PLS model, artificial neural network
models were trained for comparison.
In order to determine the optimal configurations for the ANN
model, parameters were optimized while training, cross-validat-
ing, and testing the network in an effort to obtain the best predic-
tive final model. With the learning rate (LR) 0.01, the momentum
rate (MR) 0.05, the parameters that were modified were the num-
ber of nodes in the hidden layer (from 1 to 4), and the number of
total training cycles (from 10,000 to 100,000) for training the neu-
ral network models with the configuration NN11-n-1 which have
11 input nodes, a hidden layer of n nodes and one output node.
Using each set of these two variables, models were optimized
and evaluated based on the correlation of trained and observed
data (r2-value) and root-mean square deviation (rmsd) for the
training results, as well as for the leave-one-out validation results,
which generated separate correlation coefficients ðr2CVÞ and error
values (loormsd) (see method section for definitions).
From the 3-dimensional display in Figure 4, it can be seen that
the R value (r2 = R2) from the training results increases with the
number of training cycles and the rmsd decreases accordingly. It
can also be seen that the addition of a second and third node to
the hidden layer yielded models with greater correlation (R value).
The leave-one-out validation paints a clearer picture, and a clear
maximum of correlation is produced by the models with 1 hidden
layer containing 3 hidden nodes, affording an r2CV value of 0.721 at
30,000 training cycles (see Fig. 5). The error is also minimized with
this set of parameters to 0.257 loormsd. While one hidden layer
with three nodes generated the best overall leave-one-out valida-
tion correlation, the 30,000 training cycles also generated the best
rcv values with each of the other hidden node values, resulting in
the semispherical topology generated in Figure 5.
0.85
0.84
0.83
0.82
rcv
12000
0.81
0.8
20000
25000
0.79
0.78
30000
1
Cycles
40000
2
50000
3
Nodes
4
0.79-0.8
0.8-0.81
0.81-0.82
0.82-0.83
0.83-0.84
0.84-0.85
Figure 5. Optimization of rcv by variation of cycles and hidden nodes using the
leave-one-out validation of the artificial neural network model.
the results with the second PLS model using the same eleven
descriptors, the correlation of the leave-one-out validation was
only slightly better in the best ANN model, at 0.721 compared to
0.714 in the PLS model, and the loormsd was slightly decreased
in the leave-one-out method, from 0.261 in the PLS model to
0.257 in the ANN model. This suggested that the linear feature is
still dominant in the 45-compound dataset.
The experimental pIC50 values for the 45 arylidenamino-guani-
dino compounds which were generated from the screening data, as
well as those calculated by both the trained and leave-one-out
validated values from the neural network model NN11-3-1 are dis-
played in Table 4. The observed values are plotted against the
trained and LOO-validated values from the training method and
the leave-one-out method in Figure 6.
In principle, the artificial neural network model is able to dra-
matically improve upon the training set r2, validation rC2V, and their
corresponding errors when strong non-linear patterns are included
in a dataset in contrast to a linear model.32 However, comparing
The plotted theoretical values are, not surprisingly, very linear
when compared to the observed values (see Fig. 6). Each of the
points lies within about half an order of magnitude of the diagonal
line of perfect correlation. The leave-one-out data is very similar,
with most of the points lying within roughly one half of an order
of magnitude of the ideal rcv = 1 line. While this final predictive
model had several points which were somewhat different from
the values observed in the [3H]MK-801 displacement assay, it
should be remembered that this model is a tool to be used for
pre-synthetic prediction of a molecule’s activity. This model should
be able to make a good prediction for the IC50 value of any com-
pound with the general arylidenaminoguanidine structure, how-
ever, there may be some structural features which impact site
binding that have not yet been introduced into this ANN model,
and this could create anomalies with the current model. Therefore,
the subsequent addition of more molecular structures will refine
and re-train the model, making it more accurate and reducing
the number of outlying points.
0.935
0.93
0.925
0.92
R
0.915
0.91
0.905
0.9
Cycles
4
3
2
Nodes
1
3.4. Insights obtained from this modeling study
0.9-0.905
0.905-0.91
0.91-0.915
0.915-0.92
0.92-0.925
0.925-0.93
0.93-0.935
The ANN model in its current form is a successful predictive tool
with excellent correlation values. The standard proof of a model’s
Figure 4. Optimization of R by variation of cycles and nodes using the training
method of the artificial neural network model.

J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
1773
high predictive ability is considered to be an r2CV value of greater
than 0.5. The best artificial neural network model obtained in this
study possesses a correlation coefficient r2 of 0.829 and an rmsd of
0.200 as well as a cross-validation r2CV of 0.721 with loormsd of
0.257.
in water to be screened. However, the general finding can be a va-
lid consideration that testing further compounds which have in-
creased lipophilicity may improve their inhibitory properties at
the NMDA receptor.
While the predicted IC50 values generated from the ANN meth-
od are quite complicated when comparing them directly with the
descriptors, the two PLS models have given some new insights into
the structure-activity relationships of arylidenamino-guanidine
compounds with respect to their ability to displace bound
[3H]MK-801 in the NMDA receptor screening assay.
3.5. Conclusions
Using a combination of both the partial least squares (PLS) and
back-propagation artificial neural network (ANN) pattern recogni-
tion methods, several models have been developed to predict the
activity of arylidenaminoguanidine compounds as inhibitory mod-
In both models, increasing the ‘C-027’ descriptor would reduce
the predicted inhibition, increasing the predicted IC50 value. This
descriptor is a measure of the number of unsubstituted heteroat-
oms adjacent to a mono-protonated carbon, that is. those con-
tained within an aromatic ring (but without complete adjacent
substitution as in the case of compound 35). Therefore, inclusion
of pyridine, furan, or thiophene rings in the structure of analogs ap-
pears to be detrimental to the inhibitory activity of this class of
compounds. Conversely, in the first model, the descriptor nNHR
was shown to have a positive effect on inhibition; i.e. increasing
the number of substituted protonated amino groups increased
receptor inhibition. While this latter descriptor was not used in
either the second PLS model or the ANN model, these combined
data indicate that indole and pyrrole rings, which contain non-ba-
sic protonated nitrogen atoms, are not detrimental to NMDA
inhibition.
The first PLS model also demonstrated that decreasing the PSA
descriptor increases the inhibitory activity of the molecule. This
descriptor is a measure of polar surface area, and while it would
be difficult to differentiate most of these compounds based on
large differences in molecular surface area, this observation indi-
cates that increasing molecular size (and thus PSA) could be detri-
mental to the inhibitory activity of the molecules.
Finally, an increase in the descriptor mlogP predicted an in-
crease in the inhibitory activity of the arylidenaminoguanidino
compounds in the second model. This Moriguchi octanol-water
partition coefficient is a measure of the lipophilicity of the com-
pound. Thus, increasing the mlogP value in these analog structures
lowered the predicted IC50 value. The compounds with the highest
value for this descriptor included compound 9 [(2-biphenyl-carbo-
xylidenamino)-guanidine hydrochloride] as well as compounds 32
and 33 (the 2- and 4-methylated 1-naphthylidenaminoguanidine
hydrochloride compounds, respectively). This finding is interesting
in that these more hydrophobic compounds were difficult to eval-
uate in the screens, due to their low water solubility. This also
raises issues with regard to such compounds as (4-biphenylcarb-
oxyliden-amino)guanidine hydrochloride, which was too insoluble
ulators of the N-methyl-D-aspartate receptor complex. This was
done by correlating structural and physicochemical descriptors ob-
tained from computation software to the experimentally observed
[3H]MK-801 displacement ability of 45 synthesized and screened
arylidenaminoguanidine compounds.
A PLS model was first developed using a set of 27 inhibitory
compounds, and this model was used to predict the inhibitory abil-
ity of 18 new compounds. The statistical results for the generated
PLS model are r2 = 0.899, rmsd = 0.132, rC2V = 0.815, and loo-
rmsd = 0.227. 12 among 18 newly synthesized molecules have
been predicted with reasonable accuracy. The second generation
of models was built after adding the set of an additional 18 new
arylidenaminoguanidines. The statistical results for the PLS model
are r2 = 0.814, rmsd = 0.208, rC2V = 0.714, loormsd = 0.261. The de-
crease in correlation in the second generation of the PLS model is
due to the increase in sample size, in addition to the inclusion of
several biphasic compounds.
The ANN model was created utilizing the eleven descriptors
from the second PLS model for comparison. The fact that the qual-
ity of the ANN model (r2 = 0.828, rmsd = 0.200, rC2V = 0.721, loo-
rmsd = 0.257) is similar to the second PLS model indicates that
the feature between the inputs and the output is majorly linear.
For all the screened compounds in the dataset used to build mod-
els, each had the same common structural feature, that is, an aro-
matic ring attached to an iminoguanidine moiety. The majority of
the molecules screened in the laboratory and predicted in these
models have correlated well between observed and predicted
inhibitory values (see Tables 2 and 4 and Fig. 1, 3 and 6).
These computational models are vital tools in the development
of future generations of arylidenaminoguanidino compounds, as
they are able to predict inhibition of the NMDA receptor complex
by this series of compounds in silico. Arylidenamino-guanidino
compounds with low predicted IC50Pot-values (i.e., high NMDA
receptor inhibitory potency) will be synthesized and examined in
the screening assays, while those compounds with high predicted
IC50Pot-values (i.e., low NMDA receptor inhibitory potency) will
be ignored. In this way, a predictive structure-based ‘pre-screen-
ing’ of the compounds will save a substantial amount of otherwise
wasted time and cost in synthesis and screening.
Furthermore, as more compounds are synthesized and
screened, the number of samples in the dataset used for generating
new QSAR models will increase. In this case, linear regression mod-
els would meet their limitation, while the non-linear ANN models
would show their power by recognizing the additional new struc-
tural ‘knowledge’ with its flexible learning system leading to highly
predictive models. The ability of these models to continually
evolve in accuracy and complexity with the addition of further
training sets of arylidenaminoguanidino compounds makes them
invaluable in the further development of NMDA receptor inhibi-
tory modulators as potential drug candidates. In this respect, com-
pound
24,
JR-220
(4-(chloro-benzylidenamino)guanidine
hydrochloride), in this current study, has recently been tested in
a screening hierarchy for anti-relapse and neuroprotective activity,
ranging from cell-based assay, through tissue culture to animal
behavior.33 This hierarchy has been validated using drugs with
Figure 6. Comparison of observed pIC50 versus trained pIC50 as well as LOO-
validated values calculated using artificial neural network model NN11-3-1.

1774
J. R. Ring et al. / Bioorg. Med. Chem. 21 (2013) 1764–1774
23. McKay, B. E.; Lado, W. E.; Martin, L. J. Pharmacol. Biochem. Behav. 2002, 72, 551.
24. Ring, R.R. Ph.D. Dissertation, University of Kentucky Doctoral Dissertation,
UKnowledge, No. 413, 2006.
known, or potential, clinical value at these targets (i.e., acampro-
sate (N-acetylhomotaurine), memantine and topiramate). JR-220
was found to be non-toxic and showed excellent activity in every
screen, with a potency 5 to 200-fold greater than that of the
FDA-approved anti-relapse agent, acamprosate.
25. (a) Tripos discovery software package with SYBYL 7.0, Tripos Inc., 1699 South
Hanley Rd., St. Louis, Missouri, 63144, USA. (b). Frisch, M. J.; Trucks, G. W.;
Schlegel, H. B.; Scuseria, G. E.; Robb, M. A.; Cheeseman, J. R.; Montgomery, Jr., J.
A.; Vreven, T.; Kudin, K. N.; Burant, J. C.; Millam, J. M.; Iyengar, S. S.; Tomasi, J.;
Barone, V.; Mennucci, B.; Cossi, M.; Scalmani, G.; Rega, N.; Petersson, G. A.;
Nakatsuji, H.; Hada, M.; Ehara, M.; Toyota, K.; Fukuda, R.; Hasegawa, J.; Ishida,
M.; Nakajima, T.; Honda, Y.; Kitao, O.; Nakai, H.; Klene, M.; Li, X.; Knox, J. E.;
Hratchian, H. P.; Cross, J. B.; Adamo, C.; Jaramillo, J.; Gomperts, R.; Stratmann,
R. E.; Yazyev, O.; Austin, A. J.; Cammi, R.; Pomelli, C.; Ochterski, J. W.; Ayala, P.
Y.; Morokuma, K.; Voth, G. A.; Salvador, P.; Dannenberg, J. J.; Zakrzewski, V. G.;
Dapprich, S.; Daniels, A. D.; Strain, M. C.; Farkas, O.; Malick, D. K.; Rabuck, A. D.;
Raghavachari, K.; Foresman, J. B.; Ortiz, J. V.; Cui, Q.; Baboul, A. G.; Clifford, S.;
Cioslowski, J.; Stefanov, B. B.; Liu, G.; Liashenko, A.; Piskorz, P.; Komaromi, I.;
Martin, R. L.; Fox, D. J.; Keith, T.; Al-Laham, M. A.; Peng, C. Y.; Nanayakkara, A.;
Challacombe, M.; Gill, P. M. W.; Johnson, B.; Chen, W.; Wong, M. W.; Gonzalez,
C.; Pople, J. A. Gaussian 03, Revision A.1, Gaussian, Inc., Pittsburgh, PA, 2003.
26. DRAGON software version 3.0, 2003, developed by Milano Chemometrics and
QSAR Research Group (http://www.disat.nimib.it/chm/Dragon.htm).; (b)
Todeschini, R.; Consonni, V. In Handbook of Molecular Descriptors; Wiley-
VCH: Weinheim (Germany), 2002.
27. Zheng, F.; Zheng, G.; Deaciuc, A. G.; Zhan, C.-G.; Dwoskin, L. P.; Crooks, P. A.
Bioorg. Med. Chem. 2007, 15, 2975.
28. Geladi, P.; Kowalski, B. Anal. Chim. Acta 1986, 185, 1.
29. Zheng, F.; Bayram, E.; Sumithran, S. P.; Ayers, J. T.; Zhan, C.-G.; Schmitt, J. D.;
Dwoskin, L. P.; Crooks, P. A. Bioorg. Med. Chem. 2006, 14, 3017.
30. Zheng, F.; Zheng, G. R.; Deaciuc, A. G.; Zhan, C.-G.; Dwoskin, L. P.; Crooks, P. A. J.
Enzyme. Inhib. Med. Chem. 2009, 24, 157.
31. Zheng, F.; McConnell, M. J.; Zhan, C.-G.; Dwoskin, L. P.; Crooks, P. A. Bioorg. Med.
Chem. 2009, 17, 4477.
32. Skapura, D. M. Building Neural Networks; Addison-Wesley: New York, NY, 1996.
33. Barron, S.; Lewis, B.; Wellman, K.; Carter, M.; Farook, J.; Ring, J.; Rogers, D. T.;
Holley, R.; Crooks, P. A.; Littleton, J. Recent Pat. CNS Drug Discov. 2012, 7, 129.
References and notes
1. Koh, J. Y.; Choi, D. W. Brain Res. 1991, 548, 318.
2. Dubinsky, J. M. Ann. N.Y. Acad. Sci. 1993, 679, 34.
3. Nicotera, P.; Lipton, S. A. J. Cereb. Blood Flow Metab. 1999, 19, 583.
4. Meldrum, B. S.; Akbar, M. T.; Chapman, A. G. Epilepsy Res. 1999, 36, 189.
5. Choi, D. W. Neuron 1988, 1, 623.
6. Dodd, P. R. Biogerontology 2002, 3, 51.
7. Morrison, L. D.; Kish, S. J. Neurosci Lett. 1995, 197, 5.
8. Li, L.; Fan, M.; Icton, C. D.; Chen, N.; Leavitt, B. R.; Hayden, M. R.; Murphy, T. H.;
Raymond, L. A. Neurobiol. Aging 2003, 24, 1113.
9. Lovinger, D. M.; White, G.; Weight, F. F. Science 1989, 243, 1721.
10. Dodd, P. R.; Beckmann, A. M.; Davidson, M. S.; Wilce, P. A. Neurochem. Int. 2000,
37, 509.
11. Littleton, J. M.; Lovinger, D.; Liljequist, S.; Ticku, R.; Matsumoto, I.; Barron, S.
Alcohol Clin. Exp. Res. 2001, 25, 5.
12. Williams, K.; Romano, C.; Molinoff, P. B. Mol. Pharmacol. 1989, 36, 575.
13. Sacaan, A. I.; Johnson, K. M. J. Pharmacol. Exp. Ther. 1990, 255, 1060.
14. Reynolds, I. J. J. Pharmacol. Exp. Ther. 1990, 255, 1001.
15. Li, G.; Regunathan, S.; Reis, D. S. J. Neurochem. 1996, 67, 1761.
16. Sastre, M.; Regunathan, S.; Galea, E.; Reis, D. J. J. Neurochem. 1996, 67, 1761.
17. Reis, D. S.; Regunathan, S. Trends Pharmacol. Sci. 2000, 21, 187.
18. Zhu, M. Y.; Iyo, A.; Piletz, J. E. Biochem. Biophys. Acta 2004, 1670, 156.
19. Iyo, A.; Zhu, M. Y.; Ordway, G. A.; Regunathan, S. J. Neurobiochem. 2006, 96,
1042.
20. Gibson, D. A.; Harris, B. R.; Rogers, D. T.; Littleton, J. M. Brain Res. 2002, 952, 71.
21. Piletz, J. E.; May, P. J.; Wang, G.; Zhu, H. Ann N.Y. Acad. Sci. 2003, 1009, 64.
22. Raasch, W.; Schafer, U.; Chun, J.; Dominiak, P. J. Pharmacol. 2001, 133, 755.