Combining QSAR classification models for predictive modeling of human monoamine oxidase inhibitors

Article abstract of DOI:10.1016/j.ejmech.2012.10.035

Due to their role in the metabolism of monoamine neurotransmitters, MAO-A and MAO-B present a significant pharmacological interest. For instance the inhibitors of human MAO-B are considered useful tools for the treatment of Parkinson Disease. Therefore, the rational design and synthesis of new MAOs inhibitors is considered of great importance for the development of new and more effective treatments of Parkinson Disease. In this work, Quantitative Structure Activity Relationships (QSAR) has been developed to predict the human MAO inhibitory activity and selectivity. The first step was the selection of a suitable dataset of heterocyclic compounds that include chromones, coumarins, chalcones, thiazolylhydrazones, etc. These compounds were previously synthesized in one of our laboratories, or elsewhere, and their activities measured by the same assays and for the same laboratory staff. Applying linear discriminant analysis to data derived from a variety of molecular representations and feature selection algorithms, reliable QSAR models were built which could be used to predict for test compounds the inhibitory activity and selectivity toward human MAO. This work also showed how several QSAR models can be combined to make better predictions. The final models exhibit significant statistics, interpretability, as well as displaying predictive power on an external validation set made up of chromone derivatives with unknown activity (that are being reported here for first time) synthesized by our group, and coumarins recently reported in the literature.

Full text of DOI:10.1016/j.ejmech.2012.10.035

Accepted Manuscript
Combining QSAR classification models for predictive modeling of human monoamine
oxidase inhibitors
Aliuska Morales Helguera, Alfonso Pérez-Garrido, Alexandra Gaspar, Joana Reis,
Fernando Cagide, Dolores Vina, M.Natália D.S. Cordeiro, Fernanda Borges
PII:
S0223-5234(12)00639-3
10.1016/j.ejmech.2012.10.035
EJMECH 5804
DOI:
Reference:
To appear in: European Journal of Medicinal Chemistry
Received Date: 26 June 2012
Revised Date: 16 October 2012
Accepted Date: 18 October 2012
Please cite this article as: A.M. Helguera, A. Pérez-Garrido, A. Gaspar, J. Reis, F. Cagide, D. Vina,
M.N.D.S. Cordeiro, F. Borges , Combining QSAR classification models for predictive modeling of
human monoamine oxidase inhibitors, European Journal of Medicinal Chemistry (2012), doi: 10.1016/
j.ejmech.2012.10.035.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

ACCEPTED MANUSCRIPT
•
•
•
•
New QSAR model to speed up the development of human MAO inhibitors
Development of a novel dataset for modeling hMAO inhibitory activity
Development of combined QSAR classification models
Improvement of model predictive capacity

ACCEPTED MANUSCRIPT
Combining QSAR classification models for
predictive modeling of human monoamine
oxidase inhibitors
a,b,c,*
d,e
a
Aliuska Morales Helguera
, Alfonso Pérez-Garrido , Alexandra Gaspar , Joana
a
a
f
g
Reis , Fernando Cagide , Dolores Vina , M. Natália D. S. Cordeiro , Fernanda Borges
a,
*
a
CIQ, Departamento de Química e Bioquímica, Faculdade de Ciências, 4169-007,
Universidade do Porto, Porto, Portugal.
b
c
CBQ / Departamento de Química, Universidad Central Marta Abreu de Las Villas,
Santa Clara, 54830, Villa Clara, Cuba.
d
e
Cátedra de Ingeniería y Toxicología Ambiental/ Departamento de Tecnología de la
Alimentación y de la Nutrición. Universidad Católica San Antonio, Guadalupe, Murcia,
España.
f
Departamento de Farmacología, Facultad de Farmacia, Universidad de Santiago de
Compostela, 15782 Santiago de Compostela, España.
g
REQUIMTE, Departamento de Química e Bioquímica, Faculdade de Ciências, 4169-
0
*
2
07, Universidade do Porto, Porto, Portugal.
To whom correspondence should be addressed. Phone: +351 220402560. Fax: +351
20402659. E-mail: fborges@fc.up.pt, E-mail; aliuskamhelguera@yahoo.es
ABSTRACT
1

ACCEPTED MANUSCRIPT
Due to their role in the metabolism of monoamine neurotransmitters, MAO-A and
MAO-B present a significant pharmacological interest. For instance the inhibitors of
human MAO-B are considered useful tools for the treatment of Parkinson Disease.
Therefore, the rational design and synthesis of new MAOs inhibitors is considered of
great importance for the development of new and more effective treatments of
Parkinson Disease. In this work, Quantitative Structure Activity Relationships (QSAR)
has been developed to predict the human MAO inhibitory activity and selectivity. The
first step was the selection of a suitable dataset of heterocyclic compounds that include
chromones, coumarins, chalcones, thiazolylhydrazones, etc. These compounds were
previously synthesized in one of our laboratories, or elsewhere, and their activities
measured by the same assays and for the same laboratory staff. Applying linear
discriminant analysis to data derived from a variety of molecular representations and
feature selection algorithms, reliable QSAR models were built which could be used to
predict for test compounds the inhibitory activity and selectivity towards human MAO.
This work also showed how several QSAR models can be combined to make better
predictions. The final models exhibit significant statistics, interpretability, as well as
displaying predictive power on an external validation set made up of chromone
derivatives with unknown activity (that are being reported here for first time)
synthesized by our group, and coumarins recently reported in the literature.
Keywords: monoamine oxidase inhibitor, selectivity, hMAO-A, hMAO-B, QSAR,
consensus
2

ACCEPTED MANUSCRIPT
INTRODUCTION
The monoamine oxidase (EC 1.4.3.4; MAO) is a flavoprotein localized in the outer
mitochondrial membrane and present in practically all mammalian tissues. The primary
role of MAO lies in the metabolism of amines and in the regulation of neurotransmitter
levels and intracellular amine stores[1].
All mammals contain two MAO isoforms, MAO-A and MAO-B. They can be
distinguished by their respective substrate preferences, by their sensitivities to the
acetylenic inhibitors clorgyline and L-deprenyl (selegiline), and by their tissue
distribution [1-3]. MAO-A, predominant isoform in the gastrointestinal tract [2], is
primarily responsible for the oxidation of tyramine. Its peripheral inhibition has been
associated with the risk for an acute hypertensive syndrome known as the “cheese
reaction”[4]. Whereas the MAO-B is the predominant subtype in the human brain,
where it acts in the breakdown of the dopamine (DA), as well as in the deamination of
β-phenylethylamine, an endogenous amine that stimulates the release of DA and inhibits
its neuronal reuptake [5]. Expression levels of MAO-B in neuronal tissue enhance 4-
fold with aging, especially in glial cells, resulting in an increased level of DA
metabolism and in the production of higher levels of dopanal and hydrogen peroxide
2 2 2 2
(H O ). Increased levels of cellular H O promote apoptotic signaling events resulting
in a decreased level of DA-producing cells [6], which are thought to play a major role in
the etiology of neurodegenerative diseases such as Parkinson’s and Alzheimer’s [6, 7].
Due to their role in the metabolism of monoamine neurotransmitters, MAO inhibitors
can be useful in treatment of psychiatric and neurological diseases. Actually, MAO-A
selective inhibitors act as antidepressant and antianxiety agents while selective MAO-B
inhibitors are an attractive therapeutic intervention for patients with Parkinson’s disease
[
1, 8-10].
The recognition of the importance of MAO as a drug target has produced a growing
interest in the development of MAO inhibitors. Privileged structures, such as indoles,
arylpiperazines, biphenyls and benzopyranes are currently ascribed as supportive
approaches in drug discovery. Different families of heterocycles, such as oxadiazolones,
tetrazoles, oxadiazinones, indenopyridazines, pyrazoles, xanthones, coumarins and their
3

ACCEPTED MANUSCRIPT
precursors (chalcones) have also been extensively used as scaffolds in medicinal
chemistry programs, namely for the search of novel MAO inhibitors [11-19].
Recently, a discovery and development of new chemical entities as human MAO
(
hMAO) inhibitors belonging to chromone (benzo-γ-pyrone) nucleus [20-23] validating
the scaffold for the design of potent, selective and reversible MAO inhibitors. In
general, chromones substituted in position-3 of γ-pyrone nucleus act as hMAO-B
selective inhibitors and those substituted in position-2 of the pyrone nucleus inhibit the
hMAO-A or are inactive. The –COOH substituent in position 3- of this heterocyclic
scaffold as well as the introduction of a chloro substituent in para position of the
exocyclic aromatic ring of chromone 3-phenylcarboxamides improve the potency and
selectivity of hMAO-B inhibitors (see Figure 1).
Figure 1 comes about here.
In spite of the considerable advancements in the studies on the interactions of MAO
isoforms with their specific substrates or inhibitors [24], there are not any available
rules for the rational design of new potent and selective MAO inhibitors. Although
several medicinal chemistry approaches can be used for the identification of hits,
generation of leads, and optimization of leads into drug candidates, no rational
guidelines have been available yet. Quantitative structure-activity relationships (QSAR)
are important strategies that can be applied for the successful design of potent and
selective MAO inhibitors [25].
QSAR methodology consists of a representation of the chemical structure using
molecular descriptors and a multivariate analysis that relates the biological activity of
one compound to its chemical structure. Several methodologies have established QSAR
to predict the inhibition on MAOs. The most of them are based on specific kind of
compound (congeneric series), e.g. coumarin, pirindole, xanthone, etc. [11, 26, 27] or
use chemicals assayed by different experimental protocols [28].
One of the first QSAR studies to predict inhibitory MAO activity is developed in 1974
by Martin et. al. [29]. The study involved a relatively small data set (20 chemicals) of
4

ACCEPTED MANUSCRIPT
aminotetralins and aminoindans with antidepressant activity reported. Linear
discriminant analysis was used as multivariate analysis technique and only two
molecular descriptors: Taft steric parameter and an indicator variable were statistically
significant. Also field-based 3D-QSAR like comparative molecular field analysis
(CoMFA) is among the computational methods that have been extensively used to this
end [24, 26, 27, 30-33]. This popular 3D QSAR approach makes suggestion regarding
steric or electronic modifications of the training set compounds that are likely to
increase their activity. Some examples on the use of CoMFA to understand MAO-
selectivity include the work of Gnerre et. al. to unveil structural requirement for
selective binding of 7-X-benzyloxy meta-substituted 3,4-dimethylcoumarin derivatives
to the MAO [27]. The chemical interpretation of the resulting CoMFA model suggests
lipophilicity as the property that should influence MAO-B versus MAO-A selectivity,
while the presence of an electron-withdrawing substituent should favor the MAO-A
inhibition. However in CoMFA, the alignment of the molecules is a main issue because
the relative interaction energies strongly depend on the relative molecular position [34].
This problem is considered [35] a leading cause to incorrect QSAR models, specifically
those created with CoMFA.
An alternative to the CoMFA model of MAO inhibitors is the use of the receptor based
alignment, recently proposed by Chimenti et. al [24]. This computational approach is
possible because the crystallographic structure of human MAO-B and MAO-A have
been recently published [36, 37]. Combining docking with CoMFA/QSAR may enhance
the probability of finding potent inhibitors. However the current methodology has
limitations like as: the high time-consuming of molecular geometry optimization and its
application only to homogeneous series of compounds due to the requirement for
structural alignment.
Other QSAR model to predict MAO-A inhibitory activity was developed by Núñez
et.al. [11] using a data set of 42 xanthones derivatives and E-state index, molecular
connectivity and shape descriptors. After the exclusion of 7 outliers the model had a
determination coefficient of 0.847. The only validation methods of the model were,
leave one out and jackknife technique. There was no division of the data set into
training and test sets. Consequently, the model external predictivity power was not
evaluated. In light of recent studies is important to make the external validation [38],
5

ACCEPTED MANUSCRIPT
this omission compromises the application of this model to predict accurately the MAO
inhibition of compounds that were not used for the model development. Some other
authors validate their MAO-QSAR model using one or two compounds that were not
used in the model development [39], and still claim their models are predictive.
On the other hand, the information on the inhibitory MAO activity of many molecules
reported in the literature is limited and directed to behavioral tests and assays on bovine
or rat mitochondrial MAOs [24]. Therefore, many of the QSAR studies to predict
inhibitory MAO activity were based on databases assayed in non-human MAO tests [11,
2
4, 26, 27, 30]. However the access to purified recombinant human MAO-A [40] and
MAO-B [41] as well as conditions for the crystallization of MAO-B complexes with
several inhibitors [36, 42], has resulted in explosive growth of amount of human MAO
inhibition data [13, 14, 17, 28, 32, 43-55]. This progress has provided new opportunities
for the design and fast computer-assisted development of drugs involved in the human
MAO metabolic pathway, since accumulating evidence has shown that inhibition data
from rat MAO test often strongly differ from data collected with similar methods from
human enzymes [56].
In the present paper, we assemble a set of over 450 different nitrogen and oxygen
heterocycles as well as other synthetic analogues obtained in our laboratories [20, 21,
2
3] or elsewhere [13, 14, 17, 28, 32, 43-55] that were evaluated towards human MAO
in a single and consistent inhibition assay [57]. Herein, 0D, 1D and 2D and molecular
descriptors along with linear discriminant analysis and feature selection algorithms were
used to build reliable predictive QSAR models from the available data. The final models
were reliable displaying significant statistics and were predictive on an external
validation set comprising chromone derivatives of unknown activity and coumarins
recently reported in the literature [58]. Our work also highlights that different
combinations of the QSAR models increase our chance for success, whereas a
conventional QSAR approach using only one model and one type of descriptors has a
higher chance to fail.
MATERIALS AND METHODS
Data set
6

ACCEPTED MANUSCRIPT
Different types of heterocyclic such as chromones, homoisoflavonoids, coumarins and
their precursors (chalcones) and a 2-hydrazinylthiazoles and pyrazoles have been
included in our database. They were experimentally assayed for their inhibitory effects
(IC50) on the hMAO-A and hMAO-B isoforms. The experiments were all conducted at
the Department of Pharmacology of Faculty of Pharmacy of University of Santiago de
Compostela in Spain, following the same in vitro assay protocols that were previously
described [57]. To note that the biological data used along the work has been measured
by a single protocol at the same laboratory.
The compounds were first classified into four groups according to their IC50 values and
the selectivity indexes (SI). The first group, designated as inhibitors, include all
chemicals with IC50 < 20 µM (pIC50 > 4.70). The second one, named as non-inhibitors,
include those compounds with IC50 > 20 µM (pIC50 < 4.70). The third group, named as
hMAO-B selective ligands, contain those substances with pSI > 1.70 while fourth one,
named as hMAO-B non-selective ligand comprise those compounds with pSI < 1.70.
pSI is the selectivity to hMAO-B and it is calculated as follow: pSI = pCI50(hMAO-B) –
pIC (hMAO-A). In this dataset, there are stereoisomers, which cannot be distinguished
5
0
by the present 2D descriptors but had nevertheless similar IC50 values; in such cases,
one of the isomers was discarded. Consequently, our data set is formed by 449 organic
compounds. Table S1 of the Supplementary Material displays a complete listing of the
compounds and the reported experimental data.
The present large data set was divided into training and test sets to obtain validated
QSAR models. In this work, the k-Means Cluster Analysis (k-MCA) technique was
applied to setup both sets representative of the entire experimental universe.
k-Means Cluster Analysis
The Joining Tree Clustering (JTC) technique was applied to the whole date set of
compounds to set the appropriate number of clusters and the Ward’s method was used as
a criterion of linking. JTC is implemented in the STATISTICA software (version 8.0)
[
59]. The distances between clusters are determined by the greatest distance between
any two objects in the different clusters (this is known as “furthest neighbors”), while
7

ACCEPTED MANUSCRIPT
the distances between cases are computed by Euclidean or City-block (Manhattan)
distances.
An appropriate number of clusters for each class; inhibitors and non-inhibitors from
three dataset (both isoforms and selectivity) were selected (Figure S1 of the
Supplementary Material). For a better visualization, the optimal number of clusters was
highlighted with a red line cutting the branches. Once the number of clusters to be
developed was estimated, the k-MCA was applied to design the training and prediction
sets. By using this technique, it is possible to obtain exactly k different clusters of
greatest possible distinction that can be assessed by comparing the means of the
resulting k-clusters. Another indicator of the good performance of this technique is the
magnitude of the Fisher ratio values from the analysis of variance developed on each
dimension (Table S3 of the Supplementary Material). For this aim, the STATISTICA
software (version 8.0) [59] was used.
After the partition of the whole data into different statistically representative cluster of
compounds, the training and test sets were selected from the members of these clusters.
Choosing the suitable variables to be used in the cluster strategy can be a problem in
this case due to the availability of a large number of molecular descriptors to undergo
subsequent variable selection while developing QSAR predictive models. However, this
problem can be solved by using a linear dimensionality reduction technique. In that way
useful structural information can be kept as much as possible to be used in the clustering
technique without rejecting any descriptor in advance. Principal Component Analysis
(PCA) has been used to identify orthogonal directions of maximum variance in the
original data projecting it into a lower-dimensionality space formed by a sub-set of the
highest-variance components. As a result, the most significant factors produced by the
PCA can be the input of the clustering analysis.
Molecular Descriptors
Different sets of 0D, 1D, 2D and molecular descriptors available in the DRAGON
(version 5.4) [60], MOE (version 2008.10) [61] and MODESLAB (version 1.5) [62]
software has been used in the present work. All them have had a long history in
structure-activity and structure-property correlation [25, 63-65]. They include, for
instance, pure topological descriptors, walk and path counts, connectivity indices,
8

ACCEPTED MANUSCRIPT
information indices, or 2D-autocorrelations. Taking into account the structural diversity
of the compounds an initial subset of descriptors was computed for each molecule from
the SMILES (Simplified Molecular Input Line Entry Specification) inputting of
chemical structures. By disregarding descriptors with constant or near constant values
inside each class, three final subset of 360, 174 and 226 molecular descriptors were
generated by using DRAGON, MOE and MODESLAB, respectively.
Modeling Technique
Linear Discriminant Analysis (LDA), specifically the LDA technique implemented in
the STATISTICA software (version 8.0) [59], was used to find the classification models
(
Eq. 1), that best describe the inhibitory activity P, as a linear combination of the
n
predictor X-variables (descriptors) weighted by the a coefficients :
P=a0+ a1X1+a2X2+.…+anXn
Eq. 1
LDA is based on finding the linear combinations with large ratios of between-group to
within-group sums of squares [66]. This technique constructs a separating hyperplane
between the two groups: e.g. inhibitors and non-inhibitors. The hyperplane is described
by the linear discriminant function (Eq. 1), which is defined by the geometric means
between the centroids (i.e. the centres of gravity) of the both groups.
In developing the models, P values of 1 and −1 were assigned to inhibitor (or selective)
and non-inhibitor (non-selective) group, respectively. However, posteriori probabilities
were used instead to assert the models classification of compounds.
Feature Selection
The Replacement Method (RM) [67] was used as the variable selection procedure,
which was implemented in the STATISTICA Visual Basic [59]. Following this method,
9

ACCEPTED MANUSCRIPT
we observed the variation in the performance of the selected model by minimizing the
ROC (Receiver Operating Characteristic) graph Euclidean Distance (ROCED) corrected
with FIT (λ) (named here as ROCFIT). The mathematical definition of the ROCFIT
parameter has been previously described [68], so we will limit ourselves to a brief
outline highlighting only its most important aspects.
By analogy with the FIT Kubinyi function [69, 70] used in regression analysis, a similar
function has been defined, named FIT (λ), which can be used in the discriminant
analysis. FIT (λ) is defined as follows:
FITλ=(1-λ)(n-l-1)(n+l2)λ
Eq. 2
where n is the number of cases, l is the number of parameters of our model and
λ is the
Wilk’s lambda. Therefore, a higher value for this parameter means improving the
usefulness of our model.
Other parameter, ROCED, has been recently introduced [68]. ROCED is based on the
ROC graph, and is useful for organizing classifiers. In a ROC graph, true positives (TP
or sensitivity) are plotted on the Y axis and false positives (FP or 1−specificity) are
plotted on the X axis. The point (1, 1) represents perfect classification and the line Y = X
represents random classification. One point in ROC graph is better than another if it is
to the northwest (TP is higher, FP is lower, or both) [71]. Thus, the most finely tuned
classifiers are those whose points fall in the left upper triangle as close as possible to the
corner.
Figure 2 shows the plot of the sensitivity vs. 1−specificity to one classifier and the two
points on ROC graph denote the classifications for both training and test sets. The
1 2
distances d and d are expressed as Euclidean distances and their values represent the
1
0

ACCEPTED MANUSCRIPT
distance between the real classifier and the perfect classifier, i.e. point (1, 1). Therefore
it is necessary that these two distances are as small as possible. Thus, ROCED, a
parameter relating them is defined as follows:
ROCED=(d1-d2+1)(d1+d2)(d2+1)
Eq. 3
Figure 2 comes about here.
A small ROCED value means improving the performance of our classifier. It was found
that the ROCED parameter gets a better balance between sensitivity and specificity for
both training and test sets than other indices, like the Wilk’s lambda [68]. However
when the ROCED parameter was used, some linear discriminant models showed the
lower statistical significance [68].
To solve this problem, another parameter (ROCFIT) has been defined (see Eq. 4). It
maintains the ROCED capabilities while improving the significance of the models due
to the inclusion of FIT (λ).
ROCFIT=ROCEDFIT(λ)
Eq. 4
11

ACCEPTED MANUSCRIPT
According with Eq. 4, the best classifiers are those with lower values for both ROCED
and ROCFIT and high value for FIT (λ).
In all cases, the calculations were performed for a number of variables between 3 and
1
2. For the evaluation of the model performance several indexes were calculated for
each model (Wilk’s lambda, sensibility (Se), specificity (Sp), accuracy (Ac), precision
Pr), enrichment factor (EF) [72], Matthews correlation coefficient (MCC) [73], Robust
Initial Enhancement (RIE) [74] and Boltzmann-Enhanced Discrimination of ROC
BEDROC) [75]) based on a prior probability of 0.5 and the area under the ROC curve
(
(
for both the training and test sets. Most of these indexes are based on the confusion
matrix (Table 1), e.g. Se = TP/(TP+FN); Sp = TN/(TN+FP); Ac = (FN+FP)/N, where N
=
TP + FP + TN +FN is the total of compounds in the analyzed set. All calculations
were performed with the STATISTICA software [59].
Table 1 comes about here.
The predictivity of our final models was also evaluated by using an external set of
compounds that was not used in the model setup. Validation of the final model with
compounds, which are not part of the training set, is a necessary step to ensure
generalization, and also of great relevance to future QSAR studies.
Y-Randomization Test
To ensure the robustness of a QSAR model we carried out a Y-randomozation test [76].
In this test, the Y-vector (response-variable) is randomly swapped and new models are
generated. This process is repeated several times (iterations: 300) and the average of the
parameters: sensitivity, specificity, accuracy, Wilk’s lambda, etc. for each model is
reported. It expected that the resulting classifiers should have lower statistical
parameters for training and test sets than the classifiers developed using the real Y-
response. If the opposite happens then an acceptable QSAR model cannot be obtained
for the specific modeling methods and data, e.g. independent variables of the model are
randomly associated to the response variable.
1
2

ACCEPTED MANUSCRIPT
Applicability Domain
There are several methods for assessing the applicability domain (AD) of QSAR models
77-79] but the most common one encompasses determining the leverage values for
[
each compound [80]. Leverage (h) values are calculated for both training and test sets
and are computed by equation 5.
hii=xiT(XTX)-1xi
Eq. 5
i
where x is the descriptor row-vector of the query chemical.
These values represent the distance from the chemical to the centroid of the modeled
space. A Williams plot, i.e. the plot of standardized residuals versus leverage values (h),
was then used for an immediate and simple graphical detection of both the response
outliers and structurally-influential chemicals in the model. From this plot, the ADs
were established inside a squared area within ± x standard deviations and a leverage
threshold h* (h* was generally fixed at 3p´/n, where n is the number of training
compounds and p´ the number of model parameters, whereas x = 2 or 3), lying outside
this area (vertical lines) the outliers and (horizontal lines) the influential chemicals. For
predictions, only the predicted activity for chemicals belonging to the chemical domain
of the training set should be proposed and used.
Analysis of the final models
Once the optimization of the models by using RM was concluded, the best and different
models were submitted to a consensus process. To this end, we firstly select the models
with the best performance and sort them according with the minimization of parameter
ROCFIT. Finally, a distance measure between these models was adopted to check the
similarity/diversity among the final models [69].
Consensus
1
3

ACCEPTED MANUSCRIPT
In order to improve the predictions, a series of combinations of the best and diverse
QSAR models was carried out. We combined the individual models in each of three
ways as follows [81]:
Majority Vote (Unanimity, Simple Majority and Plurality): Considering two classes, that
is to say, inhibitors/non-inhibitors or MAO-B selective/MAO-B non-selective, this
consensus pattern gives an accurate class label if all classifier are in agreement
(Unanimity), at least 50 percent of the classifiers have +1 values (Simple Majority) or
most of the classifiers (Plurality) give correct answers [81].
Weighted Majority Vote: If the classifiers in the ensemble do not display identical
accuracy, then this consensus pattern gives to the more competent classifiers more
power in making the final decision [81].
Naive Bayes: A Bayesian network consists of a structural model and a set of conditional
probabilities. Assume that all attributes (molecular descriptors) are fully independent
given the class, and then the resulting Bayesian network classifiers are called naive
Bayesian classifiers (simply NB). NB uses the following equation to estimate the
probability of a test instance x belonging to class c:
P(c|x)NB=P(c)j=imP(aj|c)
Eq. 6
th
where m is the number of attributes, a
j
is the j attribute value of x, the prior probability
|c). To construct naive Bayes classifiers, we
P(c) and the conditional probability P(a
j
only need to estimate the probability values of P(c) and P(a |c), j = 1, 2,...., m, from the
j
training data [81].
Chemistry
For the external validation of our strategy, we synthesized 10 new chromone derivatives
decorated at 2, 3 positions of the exocyclic aromatic moiety of the carboxamide
1
4

ACCEPTED MANUSCRIPT
substituent and 6 position of the chromone nucleus (450-459) and inspected their
hMAO inhibitory activities and selectivity. The structures, pharmacological profiles of
all new synthesized chromones are reported in Table 2, while the general synthetic
strategy to get the majority of the chromones is depicted in Scheme 1.
Recently Matos et. al. [58] reported the hMAO inhibitory activity of 15 new 3-
arylcoumarin derivatives substituted at 2, 3, 4 and 5 position of the exocyclic aromatic
ring and in the C6 of the coumarin scaffold (460-475). These compounds were also
included into our external validation set. Table 2 reports the structures and hMAO
inhibitory activities of these coumarins.
Table 2 and Scheme 1 come about here.
General
All chemicals used during this research were purchased from Sigma-Aldrich and were
pro analysis grade.
¹
H, and ¹³C NMR were obtained on a Bruker AMX 300 spectrometer, operating at
00.13 and 75.47 MHz, respectively, at room temperature. Chemical shifts are
3
expressed as δ(ppm) values relative to tetramethylsilane (TMS) as an internal reference;
coupling constants (J) are given in Hz. The mass spectra (MS) were obtained using a
VG AutoSpec electron impact spectrophotomer and recorded in m/z. Thin-layer-
chromatography was carried out on precoated silica gel 60 F254 plates with a layer
thickness of 0.2mm. The spots were visualized under UV detection (254nm). Normal or
flash column chromatography was performed using silica gel 60 0.2-0.5 or 0.040-0.063
mm, respectively. Solvents were partially evaporated using a Büchi Rotavapor.
The synthesis of the N-methyl-4-oxo-N-phenyl-4H-chromene-2-carboxamide (450) was
22, 24
performed according the procedure previously described by Borges et. al.
. The
structural data was in accordance to that described elsewhere [82] [83].
Synthesis of chromones 451, 452 and 457
The Fischer esterification reaction was performed following the procedure by Borges
and coworkers [84]. Accordingly, a solution of the chromone carboxylic acid (5mmol)
1
5

ACCEPTED MANUSCRIPT
in 75ml of the alcohol, containing 2ml of H SO conc., was refluxed for approximately
2
4
2
4h. After cooling the solvent was partially evaporated and the solution neutralized
using NaHCO (5%). After extraction with diethyl ether, the organic phases were
combined and washed with water. The solution was dried over Na SO anhydrous and
3
2
4
the solvent evaporated. The crude product was purified by flash column
chromatography.
Propyl-4-oxo-4H-benzopyran-2-carboxylate (451)
Column chromatographic conditions: silica gel 60 0.040-0.063 and chloroform.
1
Yield: 52%; H NMR (CDCl
3
): 1.05 (3H, t, J=7.4 Hz, CH
CH CH ), 7.13 (1H, s, H(3)), 7.46 (1H, ddd,
J=7.0, 7.7, 1.0 Hz, (H6)), 7.62 (1H, d, J=8.4 Hz, H(8)), 7.75 (1H, ddd, J=8.4, 7.0, 1.6
2 2 3
CH CH ), 1.76-1.90 (2H, m,
CH
2
CH
2
CH
3
), 4.37 (2H, t, J=6.6, CH
2
2
3
13
Hz, H(7)), 8.21 (1H, dd, J=7.7, 1.6 Hz, (H5)); C NMR(CDCl ): 10.8 (CH CH CH ),
3
2
2
3
2
2.3 (CH
2
CH
2
CH
3
), 68.9 (CH
2
CH
2
CH
3
), 115.2 C(3), 119.3 C(8), 124.9 C(4a), 126.2
CH CH ),
C(5), 126.4 C(6), 135.2 C(7), 152.7 C(8a), 154.2 C(2), 160.2 (COO CH
2
2
3
+
•
1
79.0 C(4). EM/IE: 232 (M ).
Propyl-4-oxo-4H-benzopyran-3-carboxylate (452)
Column chromatographic conditions: silica gel 60 0.040-0.063 and ethyl
acetate/dichloromethane (1:9)
1
Yield: 60%; H NMR (CDCl ): 1.04 (3H, t, J=7.4 Hz, CH CH CH ), 1.75-1.84 (2H, m,
3
2
2
3
CH
2
CH
2
CH
3
2 2 3
), 4.30 (2H, t, J=6.8 Hz, CH CH CH ), 7.39-7.47 (2H, m, H(6), H(8)), 7.68
13
(
1H, ddd, J=8.4, 7.0, 1.6 Hz, H(7)), 8.27-8.30 (1H, m, H(5)), 8.66 (1H, s, H(2)); C
NMR: 10.3 (CH CH CH ), 21.8 (CH CH CH ), 68.4 (CH CH CH ), 114.7 C(3), 118.7
C(8), 124.4 C(4a), 125.6 C(5), 125.8 C(6), 134.7 C(7), 152.2 C(8a), 155.9 C(2), 160.5
2
2
3
2
2
3
2
2
3
+
•
(
COOPr), 178.4 C(4). EM/IE: 232 (M ).
Ethyl-4-oxo-4H-benzopyran-3-carboxylate (457)
Column chromatographic conditions: silica gel 60 0.040-0.063 and petroleum
ether/ethyl acetate (5:5).
Yield:60%; NMR (CDCl ): 1.40 (3H, t, J=7.1 Hz, CH CH ), 4.36-4.44 (2H, m,
3
2
3
2 3
CH CH ), 7.30-7.51 (2H, m, H(6), H(8)), 7.71 (1H, ddd, J=8.4, 7.0, 1.6 Hz, H(7)), 8.26-
1
6

ACCEPTED MANUSCRIPT
13
8
.30 (1H, m, H(5)), 8.67 (1H, s, H(2)); C NMR (CDCl ): 14.7 (CH CH ), 61.8
3 2 3
(CH
2
CH
3
), 113.4 C(3), 116.7 C(4a), 118.6 C(8), 125.6 C(5), 126.7 C(6), 134.6 C(7),
+
•
1
52.7 C(8a), 162.2 C(2), 163.7 COOH, 179.5 C(4). EM/IE: 218 (M ).
Synthesis of 6-cloro-4-oxo-4H-benzopyran-2-carboxylic acid (453)
The reaction procedure was adapted from Hadjeri et al. [85]. Accordingly, ethyl-(6-
cloro-4-oxo-4H-benzopyran)-2-carboxylate (1,98mmol) was added to a saturated
solution of NaHCO
solution was acidified with HCl conc., diluted into water and extracted with ethyl
acetate and dried with Na SO anhydrous. After evaporation of the organic phase the
3
(10mL) and the reaction was kept at 80ºC for 3h. After cooling the
2
4
crude material was purified by recrystallization from ethyl acetate.
1
Yield (η): 60%. H NMR (MeOH): 7.15 (1H, s, H(2)), 7.57 (1H, d, J=8.9 Hz, H(8)),
13
7
.87 (1H, dd, J=8.9, 2.5 Hz, H(7)), 8.30 (1H, d, J=2.5 Hz, H(5)). C NMR (MeOH):
112.9 C(3), 119.4 C(8), 124.1 C(4a), 125.1 C(5), 126.2 C(6), 137.2 C(7), 155.9 C(8a),
•
1
63.8 C(2), 179.0 C(4). EM/IE: 226(M +1).
Synthesis of 6-chloro-4-oxo-4H-benzopyran-3-carboxylic acid (454)
A solution of NaClO 80%, (17 mmol) in water (400 mL) was added dropwise to a
2
mixture of 6-chloro-3-formyl-4-oxo-4H-benzopyran (17mmol) and sulfamic acid (22
mmol) in dichloromethane (22 mL) at 0°C for 30 min. After reaction the organic layer
was separated and the aqueous layer was extracted with CHCl
3
. The combined organic
phases were washed with water and brine, dried over Na
2
SO anhydrous and
4
evaporated. The desire compound was obtained by a recrystallization from n-hexane.
The reaction procedure was adapted from Ishizuka et al.[86].
1
Yield: 48%. H NMR (CDCl
3
): 10.37 (1H, s, COOH), 9.02 (1H, s, H(2)), 8.30 (1H, d,
13
J=2.6 Hz, H(5)), 7.81 (1H, dd, J=8.97, 2.6 Hz, H(7)), 7.63 (1H, d, J=8.97 Hz, H(8)). C
NMR (CDCl ): 113.3 C(3), 120.4 C(8), 124.2 C(4a), 125.8 C(5), 126.6 C(6), 136.5
3
•
C(7), 155.5 C(8a), 163.7 C(2), 164.2 COOH, 178.5 C(4). EM/IE: 226(M +1).
Synthesis of chromones 455 and 456
Chromones were obtained following the technique of Walenzyk et al., [87] with slight
modifications. To a suspension of the adequate 2’-hydroxyacetophenone (a or b)
1
7

ACCEPTED MANUSCRIPT
(5.86mmol) and diethyl oxalate (11.9 mmol) in EtOH (10mL) was added 10 mL of
sodium ethoxide solution (21% in ethanol). The formation of a solid occurs
instantaneously. The reactional medium was diluted with EtOH and then neutralized
with conc. HCl leading to the formation a white precipitate that was filtered-off. The
supernatant was partially evaporated, extracted with ethyl acetate and dried over
Na
2 4
SO anhydrous. The crude product was then purified by recrystallization from ethyl
acetate.
Ethyl-(6-cloro-4-oxo-4H-benzopyran)-2-carboxylate (455)
1
Yield :70%; H NMR (CDCl
3
): 1.44 (3H, t, J=7.1 Hz, CH
2
CH
3
), 4.47 (2H, m,
CH
2 3
CH ), 7.12 (1H, s, H(3)), 7.58 (1H, d, J=9.0 Hz, H(8)), 7.68 (1H, dd, J=9.0, 2.5 Hz,
1
3
H(7)), 8.15 (1H, d, J=2.5 Hz, H(5)) C NMR (CDCl
3 2 3 2 3
): 14.1 (CH CH ), 63.1 (CH CH ),
1
1
14.6 C(3), 120.5 C(8), 125.1 C(4a), 125.3 C(5), 132.0 C(6), 134.9 C(7), 152.4 C(8a),
+
•
54.2 C(2), 160.2 (COOEt), 177.2 C(4). EM/IE : 254 (M+2), 252 (M ).
Ethyl-4-oxo-4H-benzopyran-2-carboxylate (456)
1
Yield : 75%. H NMR (CDCl
3
): 1.44 (3H, t, J=7.1 Hz, COOCH
2
CH
3
), 4.48 (2H, m,
COOCH CH ), 7.14 (1H, s, H(3)), 7.46 (1H, ddd, J=8.0, 7.1, 1.0 Hz, H(6)), 7.63 (1H,
2
3
dd, J=8.3, 1.0 Hz, H(8)), 7.76 (1H, ddd, J=8.6, 7.0, 1.6 Hz, H(7)), 8.21 (1H, dd, J=8.5;
13
1
.6 Hz, H(5)). C NMR (CDCl
3 2 3 2 3
): 14.1 (CH CH ), 63.0 (CH CH ), 114.8 C(3), 119.0
C(8), 125.2 C(4a), 125.7 C(5), 125.9 C(6), 134.8 C(7), 152.0 C(8a), 155.5 C(2), 162.5
+
•
(
COOEt), 178.5 C(4). EM/IE: 218 (M ).
Synthesis of 2-Hydroxymethyl-4-oxo-4H-benzopyran (458)
To a suspension of ethyl-4-oxo-4H-benzopyran-2-carboxylate in methanol (40 mL) an
4
equimolar amount of NaBH was added dropwise. The reaction was kept with stirring
for 24 h. The solvent was evaporated and the residue was extracted with ethyl ether. The
combined organic extracts were washed with water, HCl (1M) and NaHCO 5% and
3
2 4
dried over Na SO anhydrous. After evaporation the residue was purified by flash
column chromatography (ethyl acetate/petroleum ether (5:5)), giving rise to the desire
product. The reaction procedure was adapted from Payard et al.[88].
1
Yield: 43 %. MP: 167-172ºC. H NMR: 2,95 (1H, t, J=6,7 Hz, CH
2
OH), 4,62 (2H, d,
2
J=6,3 CH OH), 6,50 (1H, s, H(3)), 7,40 (1H, ddd, J=7,9; 7,0; 1,0 Hz, H(6)), 7,42 (1H,
1
8

ACCEPTED MANUSCRIPT
d, J=8,6 Hz, H(8)), 7,66 (1H, ddd, J=8,6; 7,0; 1,6 Hz, H(7)), 8,18 (1H, dd, J=8,5; 1,6
13
2
Hz, H(5)). C NMR: 61,534 (CH OH), 108,4 C(3), 117,9 C(8), 123,9 C(4a), 125,2
+
•
C(6), 125,8 C(5), 133,8 C(7), 167,8 C(2), 156,2 C(8a), 178,5 C(4). EM/IE: 176 (M )
Synthesis of 6-chloro-3-formyl-4-oxo-4H-benzopyran (c)
A solution of 2-hydroxy-5-chloroacetophenone (40 mmol) in anhydrous N,N-
3
dimethylformamide (80 mL) was stirred at -10 ºC for 30 min. Then POCl (80 mmol)
was added dropwise at -10 ºC during 1h. The mixture was stirred at room temperature
for 12 h and poured into water (200 mL). After it was extracted with dichloromethane
2 4
and dried over Na SO anhydrous. After evaporation of the filtrate a yellow solid was
obtained which was recrystallized from n-hexane. The reaction procedure was adapted
from Zhao et al. [89]
1
Yield:75%. NMR H (CDCl
3
): 7.52 (1H, d, J=8.9 Hz, H(8)), 7.71 (1H, dd, J=8.9, 2.6
Hz, H(7)), 8.26 (1H, d, J=2.6 Hz, H(5)), 8.55 (1H, s, H(2)), 10.37 (1H, s, CHO); NMR
13
C: 120.7 C(4a), 120.8 C(8), 126.1 C(5), 126.7 C(3), 133.3 C(6), 135.5 C(7), 154.9
•
C(8a), 161.2 C(2), 175.3 C(4), 188.6 (CHO). EM/IE m/z: 208 (M +1).
Synthesis of 3-Hydroxymethyl-4-oxo-4H-benzopyran (459)
2 3
A suspension of formylchromone (5.6 mmol) and basic Al O (20g) in 2-propanol (200
mL) was kept, with stirring, at 75ºC for 4h. The mixture was filtered by celite and the
solvent evaporated. The crude material was purified by flash cromatography (ethyl
acetate/petroleum ether (5:5) with gradient elution) and recrystallized from ethanol. The
reaction procedure was adapted from Maturana et al.[90]
1
Yield: 48%. H NMR (CDCl
3
): 3.09 (1H, t, J=6.2 Hz CH
2
OH), 4.60 (2H, d, J=5.6 Hz
CH
2
OH), 7.96 (1H, s, H(2)), 7.44 (1H, ddd, J=7.9, 7.6, 0.9 Hz, H(6)), 7.48 (1H, d, J=8.4
Hz, H(8)), 7.70 (1H, ddd, J=8.5, 7.0, 1.6 Hz, H(7)), 8.23 (1H, dd, J=8.0, 1.6 Hz, H(5)).
13
3 2
C NMR (CDCl ): 58.7 (CH OH), 118.3 C(8), 123.1 C(4a), 123.8 C(3), 125.3 C(6),
+•
1
25.6 C(5) 134.0 C(7), 152.7 C(2), 156.6 C(8a), 178.6 C(4). EM/IE: 176 (M ).
Determination of human MAO isoform activity and selectivity
1
9

ACCEPTED MANUSCRIPT
The evaluations of the MAO inhibitory activity and calculation of the selectivity profile
were performed using the experimental protocol previously described [57].These assays
were developed at Department of Pharmacology of the Faculty of Pharmacy of
University of Santiago de Compostela in Spain.
RESULTS AND DISCUSSION
The starting group of molecular descriptors stemming from the Dragon software for the
further development of the QSAR study encompassed a total of 450 descriptors. To
reduce the number of dimensions to take into account for the cluster analysis without
losing much structural information, a PCA, implemented into DRAGON software [60],
was carried out.
In doing so, 73 principal components (PC) were calculated in the PCA. However,
judging from the magnitude (and significance p-levels) of the F values, the first
principal components of constitutional descriptors (PC01-01), topological descriptors
(
PC02-01), walk and path counts (PC03-01), connectivity indices (PC04-01),
information indices (PC05-01), 2D-autocorrelation (PC-06-01), Burden eigenvalues
PC08-01), eigenvalue-based indices (PC10-01) are the major criteria for assigning
(
cases to clusters (see Table S2 of the Supplementary Material). Setting-up these
conditions, the k-MCA and JTC techniques yielded different clusters for each subset
(Table S2 of the Supplementary Material). For the hMAO-A inhibition data, the k-MCA
splits the inhibitor compounds in five clusters comprising 90, 37, 38, 2 and 7 members
and the non-inhibitory compounds in six cluster including 78, 98, 35, 37, 27 members.
For hMAO-B inhibition data, the k-MCA splits the inhibitor compounds in six clusters
with 48, 121, 15, 75, 34 and 5 members each other, while the non-inhibitory compounds
in four clusters including 59, 34, 38 and 20 members. Finally for the selectivity data the
k-MCA splits the hMAO-B selective compounds in four clusters each one with 38, 47,
1
8
3 and 11 members and the non-selective compounds in five clusters with 65, 23, 151,
4 and 17 members.
2
0

ACCEPTED MANUSCRIPT
The selection of each test set was carried out by leaving out 20% of compounds
(specifically one chemical every four chemicals) in each cluster after the cluster was
sorted by the Euclidean distance criterion. Table 3 reports the distribution data
indicating the number of inhibitors/non-inhibitors and selective/non-selective
compounds in training and test sets, whereas Figure 3 reflects that it is possible to
obtain chemically balanced training and test sets. From a structural perspective, the
compounds inside all six data sets (three training set and three test set) are highly
diverse (Fig. 4). For example, the chemical structures of the inhibitors in the hMAO-B
training set fall into at least seven classes, such as chalcones, coumarins, chromones,
thiazoles, pyrazoles, homoisoflavonoids and flavonoid derivatives. Therefore,
considering the limited chemical diversity in the training set of previous MAO
inhibition QSAR models [11, 19, 27], this work provides a unique, large and diverse
dataset for modeling hMAO inhibitory activity. All these data were measured in one
laboratory under strict experimental control [57] and thus represent a valuable source
for making QSAR models with a broad applicability, which could be used prospectively
to predict possible inhibitory activity for diverse organic compounds.
Table 3 and Figure 3 come around here.
The first step to obtain the best predictive models derived from the training sets, by
combining LDA and RM along with the DRAGON, MOE and TOPS−MODE structural
representations, was to select the most significant variables. Towards that end, the RM
was carried out as described in the Materials and Methods section. In general, LDA
models with Wilk’s lambda (λ) and ROCFIT values for the training sets lesser of 0.74
and 3.20 respectively, were accepted. As a consequence, several models were available
for each target response. To compare them, we calculated Model distances for pairs of
models based in their molecular descriptors [69, 91]. These distances take into account
the correlation of the molecular descriptors within and between models and allow us to
cluster similar models, catching the most diverse models in such a way as to preserve
maximum information and diversity. Due to the large amount of statistical information
of this procedure, here we only show the best and different models derived from each
2
1

ACCEPTED MANUSCRIPT
molecular descriptor set and for the three target response; hMAO-A inhibition, hMAO-
B inhibition and selectivity to hMAO-B (SI) (Table 4 and Table S4 of the
Supplementary Material). The meaning of each descriptor included in such models is
shown in Table S3 of the Supplementary Material.
Table 4 comes around here.
According with the results shown in Table 4 several aspects can be summarized.
The hMAO-A models (Model 1-11) produce the best statistical models. Indices like as;
Matthews’s correlation coefficient (0.71-0.81 for the training set), accuracy (91.11% -
8
9
9
0
5.83% for the training set and 77.53% - 88.64% for the test set) sensibility (83.57% -
0.00% for the training set and 73.53% - 88.24% for the test set) specificity (86.36% -
3.18% for the training set and 83.64% - 90.91% for the test set) and precision (0.9 and
.8 for training and test sets, respectively) have the most high values while the ROCFIT
(0.34-0.90) parameters and Wilk’s lambdas (0.31-0.48) present the most low values. The
correlation matrix (Table S5 of the Supplementary Material) for each QSAR model also
indicates low correlation coefficients between the descriptors, which is desirable in the
development of QSAR models based on LDA.
A comparison of the number of models generated to hMAO-A, hMAO-B inhibitory
activities and hMAO-B selectivity showed that, the final model population for hMAO-A
training set (11 models) is higher than the hMAO-B (6 models) and selectivity (6
models) datasets.
Among three descriptor sets used in combination with the LDA classification method,
DRAGON and MOE descriptors afforded models with higher predictive power than
those using TOPS-MODE approach. The TOPS-MODE approach only gave models
with higher accuracy for the study of the hMAO-A inhibition dataset.
Results of the Y-randomization test (Table S6 in Supplementary Material) suggest that
all classification models using actual data represent robust quantitative structure-activity
correlations. Indeed, none of the models built with randomized activities of the training
2
2

ACCEPTED MANUSCRIPT
set had Ac > 60.00 % and λ < 0.97. Therefore, randomization of values of the inhibitory
activity and the selectivity produced no significant models.
External Validation for the QSAR models
In the optimization step of drug discovery the principal application of the present
developed QSAR models is the prediction of the hMAO inhibitory activity and hMAO-
B isoform selectivity of ligands that were not used in the model development. To
evaluate the prediction capability of our approach, we have considered an additional test
set (external test set). This is formed by two chemical classes, chromones and
coumarins. The first one consists of 10 novel chromone derivatives decorated at 2, 3 of
the exocyclic aromatic moiety of the carboxamide substituent and 6 position of the
chromone nucleus (450-459), which have been synthesized and biologically evaluated
here for first time (Table 2). The second one involves 15 new experimentally tested
coumarins on hMAO inhibition, which were published after these QSAR models were
developed [58].
The hMAO-A and hMAO-B inhibition data is reported in Table 2 together with the
selectivity index (SI = IC hMAO-A/IC hMAO-B). The Dragon, MOE and TOPS-
50
50
MODE molecular descriptors of these new 10 chromones and 15 coumarins have been
used as input matrix for the previously generated classifiers (1-21, Table 4). In most
cases, our QSAR models are able to assign the correct class for the external set with at
least 70% of accuracy (Table 5), being considered as predictive models. No predictive
models were also obtained, for instance Model 16. These results confirm that the
validation of the QSAR model using an external test set is an import part of the QSAR
analysis [38]. Table 5 also shows that in some models the predictive ability improves
when the AD is considered (e.g. model 15, 20, 21). The definition of the AD is useful in
the identification of chemicals that fall outside of it and for which the model does less
accurate predictions [77].
Table 5 comes about here.
2
3

ACCEPTED MANUSCRIPT
The best prediction performance is achieved by the hMAO-A models. For example,
models 2 excellently classify to the 100% of the external set as hMAO-A non-inhibitors.
By analyzing the statistical class-base parameters, good values of accuracy have been
obtained for the hMAO-B classifiers (Ac > 76%), with the exception of Model 16. The
selectivity models give adequate accuracy results. One out of five selectivity models
only gives an accuracy of 76%, the remaining ones only about 70%.
The 10 new chromones derivatives that were included in the external validation set are
an outcome of a diverse decoration of chromone scaffold in three different positions.
From the data, one can conclude that some chromone derivatives show inhibitory
activity toward hMAO-B in the micromolar range. The most active chromone 457 and
4
52 (IC50(hMAO-B) = 1.89 ± 0.08 µM and 5.65 ± 0.38 µM, respectively) have a
carboxylic acid function in the 3 position of γ-pyrone. Compound 457 also shows the
highest hMAO-B selectivity from the external set (SI > 52.91). All chromones in the
external set are inactive toward hMAO-A. From this chromone data, and in accordance
with recent published data of our group [20, 21, 23], we can conclude that the same type
of substitutions on the position-2 of γ-pyrone nucleus causes a total loss of hMAO-B
inhibitory activity (e.g. 451 vs. 452, 456 vs. 457, 458 vs. 459).
Our selectivity models are able to predict the selectivity toward hMAO-B of the
compound 457. But erroneously, most of them also recognize the chromones 452
(models: 18-21), 454 (models: 17-21) and 459 (models: 17-19) as hMAO-B selective
inhibitors (Table 2), increasing the false positive rate of these single classifiers. A
possible explanation to this model fail could be the similar structure of the false positive
ligands and other selective ones, e.g. 457 vs. 452 (Table 2). It interesting to note, that
4
52 is not selective to hMAO-B, however it is considered by our selectivity models as
hMAO-B inhibitor (models: 18-21 in Table 2).
The other 15 chemicals in the external set are the result of several substitutions in 5
different positions, explored to improve their selectivity [58]. Many of them display
high hMAO-B inhibitory activity and selectivity with IC50 values in nanomolar range. It
confirms that the coumarin nucleus is an attractive scaffold in the MAO inhibitor
development, which has been extensively explored to this end [33].
2
4

ACCEPTED MANUSCRIPT
The analysis of the experimental and predicted data of the external set containing
coumarins by the above mentioned models supports that the AD definition is essential.
For instance, all the compounds excluded from the AD of models 7, 8, 9, 20 and 21
(Tables 2 and 6) are coumarin derivatives and most of them were bad predicted by these
classifiers. Then considering the AD of the models, it is noteworthy how most of the
selective coumarins in the external set are correctly classified. Some exceptions,
regarding the quality of the potency and selectivity predictions are represented for
coumarins 468 and 473. Surprisingly, the majority of the models wrongly predicted
them as hMAO-B non-inhibitors (e.g. coumarin: 468 and models: 12, 13 and 15, 16)
and then they were perfectly predicted as selective against hMAO-B.
In conclusion, it is indispensable to consider AD of the model in order to make reliable
predictions. The comparison of all experimental data with the predicted
pharmacological activities by hMAO-A, hMAO-B and selectivity models on the
external set demonstrated in general, the good prediction performance of our individual
models. However a logical combination of these individual models could improve the
theoretical predictions.
Structural interpretation of the Inhibitor selectivity to hMAO-B
The structural interpretation of all models involved in the modeling of three endpoints is
a very hard task. So, we only analyzed descriptors with high frequency found in
validated hMAO-B selectivity models to elucidate the chemical features that may be
responsible for hMAO-B selectivity. “The selectivity calculations were performed based
on reliable experimental data (see Materials and Methods). Table 6 lists the top 11 most
frequent descriptors found in three validated and different Dragon and two MOE models
with both accuracies of training and test sets greater than 74%.
Table 6 comes around here.
2
5

ACCEPTED MANUSCRIPT
In fact, it is expected that the descriptors selected by the most validated and different
models should be critical chemical determinants of the respective biological activity.
Among all descriptors families, the contributions of shape descriptors were found to be
significant, suggesting a high importance of steric factors in the enzyme selectivity [92].
Thus, the shape-related Dragon/MOE most frequent descriptors in Table 6 are
associated with van der Waals volumes or areas, such as MATS3v, PEOE_VSA+1 and
Q_VSA_FNEG and one is related with self-returning walk count of order 5 (SRW05).
The latter is closely related with the presence of five member rings in the chemical
structure and it was found that the presence of these rings occurs predominantly in non-
selective ligands. This finding is in agreement with the previous SAR results obtained
with
a
series of N1-thiocarbamoyl-3,5-di(hetero)aryl-4,5-dihydro-(1H)-pyrazole
derivatives [52] where the simultaneous presence of heteroaromatic substituents in 3-
and 5-position of the pyrazoline ring (e.g. compounds 379, 380, 384 and 385, Table S1
of the Supplementary Material) leads to a decrease in the potency and selectivity for the
hMAO-B activity.
Table 7 comes around here
Consistently, three of the most frequent descriptors are based on the counting of atom-
centered fragments, C-019 and of chemical functional groups, nArCO and nCrs, and
bonds (b_1rotN). The first one (C-019) describes the =CRX fragment (Table 7), where
R represents any group linked through carbon; X represents any electronegative atom
(O, N, S, P, Se, halogens) and = represents a double bond. This fragment is included
mainly in non-selective ligands (Table 7) and it is in correspondence with experimental
evidences. For instance, Borges et al. found that chromones with substituents at the
position 2 of γ-pyrone nucleus (e.g. chromone76, Table 7) in general gives rise to a total
loss of activity while the same type of substitutions at the position 3 of γ-pyrone nucleus
act preferably as hMAO-B inhibitors and are hMAO-B selective [20, 23]. In addition,
Santana et al. also found that coumarins with electronegative groups substituted at the
position 3 of γ-pyrone nucleus (e.g. coumarin 102, Table 7) decrease the hMAO-B
2
6

ACCEPTED MANUSCRIPT
selectivity [28]. In contrast to previous referred descriptor, the chemical group that
represents to aromatic ketones (nArCO) can be mapped onto selective and non-selective
molecules. However, it can be found predominantly in selective ligands, especially in
compounds 260-294 (Table S1 of the Supplementary Material). These 35 chemicals
belong to the homoisoflavonoid family, of which 19 are selective ligands and 16 are
non-selective. It is interesting to note that exactly the similar behavior presents nCrs and
b_1rotN descriptors, which represent number of ring secondary C(sp3) and rotatable
single bonds, respectively. They can be found in selective and non-selective ligands.
Another molecular descriptor that was frequently present was the average of molecular
weight (AMW), Geary autocorrelation - lag 3 / weighted by atomic polarizabilities
(GAT3p) and descriptors calculated from the eigenvalues of a modified graph distance
adjacency matrix, weighted with partial charges (GCUT_PEOE_2).
The results of more frequent descriptors analysis reinforce our knowledge of the
subtype selectivity and may potentially provide important hints useful in the design of
hMAO-B selective inhibitors. In addition to that, these molecular descriptors were
derived from molecular structure using low computational resources, making them very
attractive and widely used in Medicinal Chemistry [25].
Consensus Predictions of Inhibition Activity (pIC50) and hMAO-B Selectivity for
Compounds Used as Test Sets and External Validation set for the hMAO Ligand
Data Sets.
Even though the predictive ability of individual models was, in general, good, we have
applied the model consensus in an attempt to improve the accuracy on the predictions.
By this method, the best classifiers (Table 4) were combined to create more useful
multi-classifiers. The multi-classifier accuracy achieved by using different subsets of
individual classifiers and different output combination (majority vote, weighted
majority vote and Naive Bayes) are shown in Table 8. To evaluate the prediction
capability of these combinatorial QSAR approaches, the test set and an additional test
set (external set, Table 2), consisting of 10 novel chromone derivatives (450-459) and
1
5 coumarins (460-475) were considered.
2
7