Rational Use of Heterogeneous Data in Quantitative Structure-Activity Relationship (QSAR) Modeling of Cyclooxygenase/Lipoxygenase Inhibitors

Article abstract of DOI:10.1021/acs.jcim.8b00617

Numerous studies have been published in recent years with acceptable quantitative structure-activity relationship (QSAR) modeling based on heterogeneous data. In many cases, the training sets for QSAR modeling were constructed from compounds tested by dif

Full text of DOI:10.1021/acs.jcim.8b00617

Subscriber access provided by Karolinska Institutet, University Library
Computational Chemistry
Rational Use of Heterogeneous Data in QSAR
Modeling of Cyclooxygenase/Lipoxygenase Inhibitors
Alexey A. Lagunin, Athina Geronikaki, Phaedra Eleftheriou, Pavel Pogodin, and Alexey V. Zakharov
J. Chem. Inf. Model., Just Accepted Manuscript • Publication Date (Web): 28 Jan 2019
Downloaded from http://pubs.acs.org on January 28, 2019
Just Accepted
“Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted
online prior to technical editing, formatting for publication and author proofing. The American Chemical
Society provides “Just Accepted” as a service to the research community to expedite the dissemination
of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in
full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully
peer reviewed, but should not be considered the official version of record. They are citable by the
Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore,
the “Just Accepted” Web site may not include all articles that will be published in the journal. After
a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web
site and published as an ASAP article. Note that technical editing may introduce minor changes
to the manuscript text and/or graphics which could affect content, and all legal disclaimers and
ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or
consequences arising from the use of information contained in these “Just Accepted” manuscripts.
is published by the American Chemical Society. 1155 Sixteenth Street N.W.,
Washington, DC 20036
Published by American Chemical Society. Copyright © American Chemical Society.
However, no copyright claim is made to original U.S. Government works, or works
produced by employees of any Commonwealth realm Crown government in the course
of their duties.

Page 1 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
Rational Use of Heterogeneous Data in QSAR
Modeling of Cyclooxygenase/Lipoxygenase
Inhibitors
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1,2
3
4
2
Alexey A. Lagunin* , Athina Geronikaki , Phaedra Eleftheriou , Pavel V. Pogodin , Alexey V.
5
Zakharov
1
Pirogov Russian National Research Medical University, Ostrovitianov str. 1, Moscow, 117997,
Russia;
2
Institute of Biomedical Chemistry, Pogodinskaya Str., 10/8, Moscow, 119121, Russia;
3
School of Pharmacy, Aristotle University, Thessaloniki, 54124, Greece;
4
School of Health and Medical Care, Alexander Technological Educational Institute of
Thessaloniki, Thessaloniki, 57400, Greece;
5
National Center for Advancing Translational Sciences (NCATS), National Institutes of Health,
Rockville, MD 20850, USA
KEYWORDS. Heterogeneous data, QSAR, cyclooxygenase, lipoxygenase, thiazolidinone
derivatives, IC50.
ACS Paragon Plus Environment
1

Journal of Chemical Information and Modeling
Page 2 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
ABSTRACT
Numerous studies have been published in recent years with acceptable quantitative QSAR
modelling based on heterogeneous data. In many cases, the training sets for QSAR modelling
were constructed from compounds tested by different biological assays, contradicting the opinion
that QSAR modelling should be based on the data measured by a single protocol. We attempted
to develop approaches that help to determine how heterogeneous data should be used for the
creation of QSAR models on the basis of different sets of compounds tested by different
experimental methods for the same target and the same end-point. To this end, more than one
hundred QSAR models for the IC values of ligands interacting with cyclooxygenase 1,2 (COX)
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
5
0
and seed lipoxygenase (LOX), obtained from ChEMBL database, were created using the
GUSAR software. The QSAR models were tested on the external set, including 26 new
thiazolidinone derivatives, which were experimentally tested for COX-1,2/LOX inhibition. The
derivatives’ IC values varied from 89 to 26 µM for LOX, from 200 to 0.018 µM for COX-1
5
0
and from 210 to 1 µM for COX-2. This study showed that the accuracy of the models depends on
the distribution of IC values of low activity compounds in the training sets. In the most cases,
5
0
QSAR models created based on the combined training sets had advantages in comparison with
QSAR models based on a single publication. We introduced a new method of combination of
quantitative data from different experimental studies based on the data of reference compounds,
which was called “scaling”.
ACS Paragon Plus Environment
2

Page 3 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
INTRODUCTION
A significant amount of experimental quantitative data regarding the biological activity of
chemical compounds has become available in recent years. These data have been collected using
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
both freely (e.g. PubChem , ChEMBL , Binding DB ) and commercially (e.g. Wombat )
available databases. The availability of these data is very attractive and valuable for use in QSAR
modelling for different tasks in medicinal chemistry (e.g., virtual screening, lead optimization,
ADMET estimation). At the same time, it is known that the quality of the data and its correct use
are the most important components of successful QSAR modelling. Several studies have
elucidated the problems of data usage in QSAR modelling. Cronin and Schultz wrote that the
“data should ideally be measured by a single protocol, ideally even by the same laboratory and
5
by the same workers” and “biological data should ideally be from well-standardized assays,
5
with a clear and unambiguous endpoint” . They consider receptor-binding assays in
5
pharmaceutical research to be good examples of high quality biological data . It is also known
6
that IC and K are the most accurate values derived from experimental pharmacological data .
5
0
i
Therefore, the IC value is a reasonable endpoint for QSAR modelling. However, when we
5
0
consider the available experimental data presented in scientific publications we can see that in
many cases there are different assays for the same target. The represented data differ in the
number of studied compounds, the distribution of IC values, and the maximal and minimal IC
5
0
50
values. In addition, data for the same targets from different organisms are usually used in
preclinical studies of drugs. This diversity of data raises questions related to the approaches for
data selection for QSAR modelling. How can one use the experimental data from different
studies with the same target and the same end-point? Should we select the data provided only by
one study or from studies that use the same assay? May the data from several assays be
ACS Paragon Plus Environment
3

Journal of Chemical Information and Modeling
Page 4 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
combined? Which is better: the creation of QSAR models based on a single study and then the
use of them to create a consensus prediction or the combination of all the data into one general
training set? The problems of data selection and representation for QSAR modelling have been
7
5
8
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
previously analysed (e.g. OECD principles , Cronin and Schultz , Dearden et al. , Fourches et
9
10
11
al. , Cherkasov et al. , Tarasova et al. ). However, these studies did not provide answers to
many of these questions. We consider the rational selection of heterogeneous experimental data
to be essential to QSAR development because many quantitative QSAR modelling studies based
on heterogeneous data have been published in the recent years (Table 1).
The aim of this article is to study different approaches for preparation of training sets to answer
the abovementioned questions based on the example of the creation of QSAR models for
inhibitors of cyclooxygenase (COX) and lipoxygenase (LOX) using the experimental data
available in ChEMBL. ChEMBL is one of the most well-known freely available web sources
containing data on structures and experimentally determined biological activities (including
interactions with targets) for more than 1.7 million compounds. We also introduced a new
method for creation of combined training sets from different experimental studies based on the
data of reference compounds, which was called “scaling”.
ACS Paragon Plus Environment
4

Page 5 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
Table 1. Examples of recently published studies with quantitative QSAR modelling based on heterogeneous data of compounds
interacting with drug targets.
Target
Unit
Method
Ntr
Ntest
Descriptors
Characteristics
Ref.
HIV-1 reverse
transcriptase
IC50
SCR, RBF
22-319
30%
6 Phys-chem. and
substructural descriptors
Different assays, median [11]
data.
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
1
1 COX from different IC50
species
Gradient-boosting
machines, RF, SVM
3228
1/6
797 PaDEL descriptors,
Morgan fingerprints
calculated by RDkit
Different assays, average [12]
data
Thyroid hormone
receptor
IC50
IC50
Linear regression
55-139
12.50%
1/6
Optimal descriptors on the Same assay
basis of SMILES attributes
[13]
[14]
DHFR rat, DHFR
homo, F7, IL4, MMP2,
CHRM3, NPY1R,
NPY2R
RF, GBM, SVM radial
344-779
Morgan fingerprints, phys- Different assays
chem. properties
calculated by RDkit
Thyroid hormone
receptor
IC50
Random Forest
QSAR:
QSAR: 13- 508 descriptors from
Different
publications [15]
1
29-181;
Docking:
5-210
21;
Docking:
12-13
Dragon
with the same assays,
averaging data or manual
curation [9].
1
positive
examples
positive
examples
ACE
IC50
IC50
Generalized linear models 245
with a Gamma distribution
18
38 phys-chem. descriptors
Dragon
different assays and [16]
species
HDAC 1, 6
kNN
95-108
15-18
Different
assays; [17]
normalization of IC50
values
P-gp
Eflux
Ratio
BA/AB,
Ki
SCR, Bayesian approach
94-256
20%
6 phys-chem. and
substructural descriptors
Different
averaging of IC values.
assays, [18]
50
ACS Paragon Plus Environment
5

Journal of Chemical Information and Modeling
Page 6 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
ACE, AChE, BDR;
COX-2; DHFR, CB2R
IC , K
ANN
67-1361
10%
FP2 and ECFP6 (1024
bits), MACCS (256 bits)
fingerprints
Different
averaging of IC values
assays; [19]
5
0
i
50
AChE
PDE 4
IC50
MIA-QSAR (Multivariate 26
analysis of images)
8 scaffolds pixels of 2D images
with
derivatives
Different assays
[20]
scaffolds
with
derivatives
corresponding to chemical
structures
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
IC50
RF, ANN, kNN
812
203
More 1000 descriptors
from Dragon
Different
exclusion of duplicate
structures
assays, [21]
HIV-1 protease
MHC I molecules
Kinases
Ki
MLR
64
35
100 phys-chem.
descriptors
Different assays
Different assays
Different assays
[22]
[23]
[24]
IC50
IC50
STR, MLR
126
43
89 phys-chem. properties
descriptors
kernel-QSAR based on 0-600
similarity of sequences and
binding sites of kinase,
GA, Tanimoto similarity
25%
FCFP_6 fingerprints
1
3
8 antitargets
0 antitargets
IC , K
50
SCR, Bayesian models
60-1366
15-344
20%
6 phys-chem. and
substructural descriptors
Different
averaging data
assays, [25]
i
i
IC , K
RBF,
models
SCR,
Bayesian 100-4061
6 phys-chem. and
substructural descriptors
Different assays, median [26]
data
5
0
N – number of compounds in the training set(s); N – number of compounds in the test set(s) ACE – angiotensin-converting enzyme; HDAC – histone
tr
test
deacetylase; AChE – acetylcholinesterase; BDR– benzodiazepine receptor; COX – cyclooxygenase; DHFR – dihydrofolate reductase; CB2R– cannabinoid
receptor subtype 2; CHRM3 - human muscarinic acetylcholine receptor M3; NPY1R, NPY2R - human neuropeptide Y receptor type 1,2 PDE 4 –
phosphodiesterase 4; MHC I – major histocompatibility class I; P-gp – P glycoprotein; F7 - human factor-7; IL4 - human interleukin-4; MMP2 - human matrix
metallopeptidase-2; SVM – Support Vector Machines; RF – Random Forest; kNN – k Nearest Neighbours; ANN – Artificial Neural Network; MLR – Multiple
linear regression; STR - stepwise regression; SCR – self-consistence regression; GA – genetic algorithm; RBF - Radial Basis Function; GBM - Gradient
Boosting Machines. Dragon - software for the calculation of more 4000 molecular descriptors; FP2, MACCS and FCFP6 – different types of fingerprints.
ACS Paragon Plus Environment
6

Page 7 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
Cyclooxygenase and lipoxygenase are well known drug targets for non-steroidal anti-
inflammatory agents. Considerable experimental data are available about the interactions of
drug-like compounds with these targets. The creation of drugs with simultaneous inhibition of
cyclooxygenase and lipoxygenase is considered to be a prospective direction in the creation of
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
27
new, effective and safe non-steroidal anti-inflammatory agents . Compounds with combined
COX-1/2 and LOX inhibition have multiple advantages as non-steroidal anti-inflammatory drugs
because they act on the two major arachidonic acid metabolic pathways and possess a wide range
28
of anti-inflammatory activities . Since leukotrienes, produced by the action of lipoxygenase,
play a role in blood coagulation and gastric tract irritation, LOX inhibition has a positive effect
in anti-inflammatory action and prevents the GI tract irritation that results from COX-1
inhibition and the prothrombotic effect resulting from COX-2 inhibition. As a result, recent
2
9
studies have focused on the development of dual-acting COX/LOX inhibitors .
1
8,25,30,31
We have experience in QSAR modelling using heterogeneous data
and in the
creation of new dual cyclooxygenase/lipoxygenase inhibitors based on the analysis of structure-
29,32
activity relationships
. This study is based on experimental data obtained during the creation
of new inhibitors of cyclooxygenase and lipoxygenase and on IC50 values in the literature
obtained from ChEMBL for ovine and human COX-1, human COX-2 and soya been
lipoxygenase.
MATERIALS AND METHODS
ChEMBL
Data on the structures and IC50 values (concentration at half-maximum inhibition) of
ligands which were experimentally tested for the inhibition of human and ovine COX-1, human
ACS Paragon Plus Environment
7

Journal of Chemical Information and Modeling
Page 8 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
COX-2 and soya been lipoxygenase (Seed lipoxygenase-1) were retrieved from ChEMBL
(version 17). To evaluate experimental errors in the data, we calculated the Standard Deviation
(SD) of the IC values of all the compounds in initial ChEMBL data for which two or more IC
5
0
50
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
values were given for the same ligand and the same drug target (COX-1, COX-2 and soya LOX)
in different experiments. This calculation was made before any transformation of data described
below.
Table 2. Standard Deviation (SD) of IC50 values of compounds tested on the studied
targets that have at least two IC values.
5
0
Target
Number of compounds Number of IC values Mean SD of log10(IC , nM)
50 50
COX-1 human
COX-2 human
LOX soya
206
710
20
632
1856
48
0.377
0.403
0.344
0.428
COX-1 ovine
42
201
The mean SD values from Table 2 show the level of errors which may be achieved by the
best QSAR models created based on the data for the studied targets.
The data on compounds tested on human COX-1 were used for creation of training and test
sets for the investigation of different approaches for the use of heterogeneous data in QSAR
modelling. During the selection of the data, the following limitations related to GUSAR were
used:
1
. The selected chemical compounds should be single organic electroneutral molecules,
with molecular weights varying from 50 Da to 1250 Da;
2. The number of compounds with estimated IC values discussed in one article should
5
0
be at least twenty.
ACS Paragon Plus Environment
8

Page 9 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
Estimation of accuracy of prediction of QSAR models
2
2
R , Q and RMSE (Root Mean Square Error) values were used for the estimation of the
2
quality and accuracy of the prediction results given by the created QSAR models. R (the
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
2
coefficient of determination) was calculated for the data used as the training sets (R ) and the
tr
2
external validation test sets (R test) using the following equation:
2
2
2
R = 1−
∑(yobs
− ypred ) /
∑(yobs − ymean )
,
where yobs is an observed dependent variable, ypred is a calculated dependent variable, and
ymean is the mean value of the dependent variable (calculated for the values of a training set for
2
2
test
2
R or for the values of a test set for R ). The R values represent the relationship between the
tr
33
2
predicted and the observed values of the measured biological activity and R values closer to 1
2
2
indicate successful predictions. Q is a cross-validated R calculated during the leave-one-out
cross-validation procedure using the data from a training set.
RMSE was calculated using the following equation:
2
∑(
yobs − ypred
)
RMSE =
n
,
where n is the number of objects.
An RMSE closer to 0 is indicative of a successful prediction.
GUSAR software
25,31,34
All QSAR models were developed using the GUSAR software
. GUSAR uses a
combination of three types of descriptors: whole-molecule descriptors, QNA (Quantitative
3
5
Neighbourhoods of Atoms) descriptors , and descriptors based on the PASS (Prediction of
ACS Paragon Plus Environment
9

Journal of Chemical Information and Modeling
Page 10 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
Activity Spectra for Substances) algorithm for predicting the biological activity spectra of
compounds based on substructural atom centric MNA (Multilevel Neighbourhoods of Atoms)
3
0
descriptors . The whole-molecule descriptors used in GUSAR are the topological length,
topological volume, lipophilicity, number of positive charges, number of negative charges,
number of hydrogen bond acceptors, number of hydrogen bond donors, number of aromatic
atoms, molecular weight, and number of halogen atoms. The QNA descriptors are defined by
two functions, P and Q. The values for P and Q for each atom i are calculated as follows:
1
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Pi = Bi (Exp(− C)) B ,
(1)
∑k
ik
k
2
1
Q = B (Exp(− C)) B A ,
(2)
i
i
∑k
ik
k
k
2
where k are all the other atoms in the molecule and
1
2
1
2
−
A = (IP + EA ), B = (IP − EA )
.
(3)
k
k
k
k
k
k
In this equation, IP is the first ionization potential, EA is the electron affinity for each atom
35
in electronvolts, and C is the connectivity matrix for the molecule . The standard values of IP
and EA of atoms in a molecule were collected from the literature. They have been published
3
5
previously in the Appendix of one of our earlier publications . Two-dimensional Chebyshev
polynomials are used for approximating the functions P and Q over all the atoms in the molecule.
3
0
The PASS biological descriptors are calculated using the PASS algorithm , which predicts
a wide range of biological outcomes including various mechanisms of action describing
interaction of ligands with targets, transporter protein binding, gene expression activities, adding
up to approximately 6400 "biological activities" at a mean prediction accuracy threshold of at
least 95%. The output from PASS is the probability, for each predicted outcome, that the
compound will be active (P ) and the probability that it will be inactive (P ). The difference
a
i
ACS Paragon Plus Environment
1
0

Page 11 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
between these two values (P -P ) was used as a molecular descriptor for the regression analysis in
a
i
GUSAR. The results of PASS prediction for all 6400 "biological activities" are used as
descriptors and the ones that show the highest correlation with the dependent variable (IC ) are
5
0
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
selected by self-consistent regression.
In GUSAR the scale of QNA and PASS based descriptors ranges from -1 to 1. So, no
additional normalization is required for these types of descriptors. Only whole-molecule
descriptors are normalized using a standard Z-score normalization procedure.
It has previously been shown that self-consistent regression (SCR) can be successfully
25,31,34,36,37
applied to different QSAR tasks
. The basic purpose of the SCR method is to remove
the variables that poorly describe the target value. The final number of variables in the QSAR
equation selected after the SCR procedure is significantly smaller compared to the initial number
3
7
of variables. In addition, it has been shown that SCR is robust against noise in the data . The
regression coefficients, obtained from SCR, reflect the contribution of each particular descriptor
(variable) to the final equation. The higher the absolute value of the coefficient, the greater its
contribution. Thus, regression coefficients obtained after SCR can be used for the weighting of
descriptors (variables) according to their importance. In this approach, the descriptors are
weighted during the calculation of the radial basis functions (RBF) by the coefficients obtained
from SCR. RBF-SCR has previously been compared to different modern machine learning
approaches using 14 data sets containing nine physicochemical properties and five toxicity
3
9
endpoints . It was shown that the RBF-SCR method provides more accurate prediction results
than other methods, even including consensus predictions of these methods.
Applicability domain estimation
ACS Paragon Plus Environment
1
1

Journal of Chemical Information and Modeling
Page 12 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
GUSAR simultaneously uses three different approaches for the estimation of the model
applicability domains: similarity, leverage, and accuracy assessment:
− Similarity. For each compound, the pair-wise distance to each of its three nearest
neighbours (3NN) in the training set is calculated using Pearson’s correlation
coefficient in the space of the independent variables obtained after SCR. The
compound is considered to be in the applicability domain of the model if the average of
these three distances is less than or equal to 0.7.
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
−
Leverage. Leverage calculations are a method for identifying outliers based on the
T
T
−1
Leverage = x (X X ) x
contribution of each molecule to its own predicted value:
,
where x is the vector of the descriptors for a test compound, and X is the matrix formed
from the rows corresponding to the descriptors of all the molecules in the training set.
A compound is considered to be outside the applicability domain of a model if its
th
leverage is higher than the 99 percentile in the distribution of the leverage values
calculated for the training set.
−
Accuracy assessment. In this approach, the applicability domain prediction for each
compound is calculated based on the Root Mean Square Error of the prediction for
three nearest neighbours in the training set, RMSE3NN, (see the Similarity metric
above) in relation to the Root Mean Square Error of prediction for the training set as a
whole (RMSEtrain): ADvalue = RMSE3NN / RMSEtrain . In this study, a threshold of 1 was
used for the ADvalue
.
Y-scrambling procedure
ACS Paragon Plus Environment
1
2

Page 13 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
GUSAR includes a Y-scrambling (Y-randomization) procedure. This procedure allows one
to ensure that the developed models do not suffer from overfitting. In this procedure the
dependent-variable vector, the Y vector (IC values), is randomly shuffled and a new QSAR
5
0
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
model is developed using the original independent-variable matrix. It is expected that the
2
resulting models should generally have low Q values. This procedure was repeated five times
2
for each model, and then the average Q value calculated.
Consensus modelling
The final predicted values for endpoints were calculated using a weighted average of the
predictions from several selected QSAR models. Each model is based on a different set of QNA
and PASS “biological” descriptors, and its predictions for each compound are weighted
according to the similarity value as calculated during the applicability domain assessment. For
each training set, the QSAR models were selected from 1600 QSAR models by internal
validation using the 5-fold cross-validation procedure in which 20% of the training set was
randomly excluded, and these data were used as an internal test set. Only models which satisfied
2
2
2
the criterion of R >0.5, Q >0.5 and R of internal cross-validation > 0.5 were selected for
consensus modelling.
Chemistry
General Procedure for the Synthesis of 2-Thiazolylimino-5-arylidene-4-
thiazolidinones
4
0
The data for compounds 1-12 were reported in our previous paper . The appropriate
arylaldehyde (6 mM) was added to a well-stirred solution of 2-(thiazol-2-ylimino)thiazolidin-4-
ACS Paragon Plus Environment
1
3

Journal of Chemical Information and Modeling
Page 14 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
one (0.8 g, 4 mM) in acetic acid (35 mL) buffered with sodium acetate (8 mM). The solution was
refluxed for 4 h and then poured into ice-cold water. The precipitate was filtered and washed
with water and the resulting crude product was purified by recrystallization from dioxane.
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Biological Assays
In the in vitro assays, each experiment was performed at least in triplicate and the standard
deviation of the absorbance was less than 10% of the average values.
Soybean Lipoxygenase Inhibition Study in Vitro. Lipoxygenase inhibition was evaluated using
32,41,42
soybean LOX type 1b as reported previously
. sLOX-1 is the soybean isoenzyme mostly
4
2
used in drug screening, although sLOX-3 has also been used . Because of the structural and
functional similarities between mammalian LOXs, sLOX is commonly used for both mechanistic
44
and inhibition studies and is widely accepted as a model for LOXs from other sources . It has
been shown that the inhibition of plant LOX activity by NSAIDs is qualitatively similar to the
4
5
inhibition of rat mast cell LOX and can be used as a simple screen for this activity . Similarity
in the inhibition behaviour between soybean LOX-1 and human recombinant 5-LOX or soybean
43,44
LOX-3 and human blood serum LOX has been observed in many cases
. The tested
compounds, dissolved in DMSO, were added to the reaction mixture at a final concentration of
1
00 µM and were preincubated for 4 min at 28 °C with soybean lipoxygenase at a concentration
-7
of 7x10 w/v. The enzymatic reaction was initiated by the addition of sodium linoleate to a final
concentration of 0.1 mM. The conversion of sodium linoleate to 13-hydroperoxylinoleic acid
was measured at 234 nm. Nordihydroguaretic acid, an appropriate standard inhibitor, was used as
the positive control (94.4% inhibition at 0.1 mM, IC = 31.3 µM).
50
COX Inhibitor Screening Assay. The COX-1 and COX-2 activities of the compounds were
measured using ovine COX-1 and human recombinant COX-2 enzymes included in the “COX
ACS Paragon Plus Environment
1
4

Page 15 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
Inhibitor Screening Assay” kit provided by Cayman (Cayman Chemical Co., Ann Arbor, MI).
2
9,32
The assay directly measures PGF2a produced by the SnCl reduction of COX-derived PGH2
.
2
The prostanoid production was quantified via enzyme immunoassay using a broadly specific
antibody that binds to all the major prostaglandin compounds. The final estimation of %
inhibition was performed at a substrate concentration much lower than the saturating
concentration. For better visualization of the differences between compounds on a 0–100%
inhibition scale, the COX-1 inhibitory activity was tested at an arachidonic acid concentration of
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
µM and the COX-2 inhibitory activity was tested at an arachidonic acid concentration of 0.1
µM. The compounds were added to the reaction mixture at a final concentration of 200 µM. IC50
values were calculated for the most active compounds. Naproxen and indomethacin, used as
positive controls, were added to the reaction mixture at the same concentration, 200 µM, as the
tested compounds.
RESULTS
QSAR modelling for compounds tested as inhibitors of human COX-1
Fourteen sets of compounds tested for the inhibition of human COX-1 were selected from
ChEMBL according to the abovementioned rules. The sets were characterized according to the
source of COX-1 isoenzyme, the products of the reaction determined to measure the enzyme
activity/inhibition and the subsequent calculation of the IC values of the tested compounds, the
5
0
number of compounds in the set, and the minimal and maximal determined IC values (Table 3).
5
0
Table 3. Datasets of compounds tested as inhibitors of human COX-1.
No
Determined Enzyme Source
Product
Nall Nsel IC50min IC50max
(nM) (nM)
log(IC50max) – Ref.
log(IC50min
)
1
-
U-937 cells
88
21
2
100000
4.7
[46]
ACS Paragon Plus Environment
1
5

Journal of Chemical Information and Modeling
Page 16 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
2
PGE2
hCOX-1
32
50
27
20
2
136000
4.8
[47]
3
4
PGE2
PGE2
hCOX-1
hCOX-1
100
5
265000
674000
3.4
5.1
[48]
[49]
*
109 66
5
6
PGE2
PGE2
U937 microsomes
24
20
22
50
5
34000
18300
2.8
3.6
[50]
[51]
hCOX-1
hCOX-1
35
96
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
7
8
PGE2
PGH2
43
23
100
20
723000
30000
3.9
3.2
[52]
[53]
human platelet-derived COX-1 25
9
TxB2
blood
34
34
4120
383360
2
[54]
1
0
TxB2
TxB2
TXB2
TXB2
TxB2
blood
blood
blood
blood
blood
28
31
39
36
29
25
21
39
31
25
1300
500
80
250
140
800000
15900
60390
49300
25610
2.8
1.5
2.9
2.3
2.3
[55]
[56]
[57]
[58]
[59]
11
12
13
1
4
Total
- Validation test set; Nall – number of all compounds studied in the publication; Nsel - number of compounds
with determined IC values that were selected for QSAR modelling.
690 451
*
5
0
As shown in Table 3, according to the data provided by the 14 selected publications, IC50
values were calculated and recorded only for 451 out of the 690 studied compounds that were
referred to in the publications. Thus, approximately 35% of the published experimental data
could not be used for the creation of quantitative (continuous) QSAR models because of the
absence of accurate IC values. The difference between the log10 values of the minimum and
5
0
maximum IC values exceeds 1.5 for all the datasets, which is one of the formal necessary
5
0
8
conditions for the creation of reasonable QSAR models . Moreover, the minimum IC values
5
0
-5
-5
are less than 10 M and the maximum IC values are higher than 10 M for all the datasets.
5
0
This is a traditional border between active and inactive compounds used in medicinal chemistry.
Therefore, we may initially expect that these data will allow the creation of helpful QSAR
models for further studies. Most of the datasets were tested using two types of assays: (1)
estimation of the PGE2 production of human recombinant COX-1 and (2) estimation of the
TxB2 level using a human whole blood assay. Thus, along with different combination
ACS Paragon Plus Environment
1
6

Page 17 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
approaches using the experimental data, we will evaluate the combination of data from sets using
different types of assays.
The dataset with the largest number of compounds (dataset #4 in Table 3) was selected as
an external validation set for the further estimation of the prediction accuracy of the created
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
49
QSAR models for the prediction of the IC values of compounds inhibiting human COX-1 .
5
0
The largest test set provided the possibility to obtain the most statistically significant results of
estimation of accuracy for the created QSAR models.
Evaluation of single QSAR models on the basis of individual datasets
We started with the creation of single QSAR models on the basis of the individual datasets
2
2
mentioned in Table 3. Acceptable QSAR models (Q > 0.5 and R of internal cross-validation >
0
.5) were obtained for only 5 of the 14 datasets (Table 4). In the second step, a prediction of IC50
2
values was made for the data from the test set and the accuracy of these predictions (R
and
test
2
RMSE ) was calculated (Table 4). Table 4 shows that in spite of the acceptable values of R ,
test
tr
2
Q , and RMSE (obtained from the internal validation of the models and comparable with the SD
tr
of the IC value (0.377) from Table 2) the values of the prediction accuracy calculated for the
5
0
2
test set were unsatisfactory, since low R
and high RMSE values were observed.
test
test
Table 4. Characteristics of single QSAR models.
2
2
2
test
№
Ref
N
R
Q
RMSEtr
R
RMSEtest
AD,%
tr
2
2
QSAR models with Q and R of internal cross-validation > 0.5
1
3
8
1
0
[55]
[48]
[53]
[57]
[59]
25
20
23
39
25
0.903
0.989
0.967
0.796
0.969
0.647
0.721
0.713
0.511
0.722
0.459
0.407
0.504
0.402
0.339
-0.553
-1.958
0.097
-0.617
0.041
1.318
2.188
1.078
1.06
100
3
89
2
100
67
14
1.05
2
2
QSAR models with Q and R of internal cross-validation < 0.5
ACS Paragon Plus Environment
1
7

Journal of Chemical Information and Modeling
Page 18 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
1
2
4
5
6
7
9
1
1
[46]
21
27
66
20
22
43
34
21
31
1.000
0.989
0.956
1.000
0.995
0.980
0.874
0.998
0.989
0.465
0.498
0.391
0.006
0.066
0.001
0.353
0.181
0.007
3.167
0.828
0.826
1.045
0.754
1.105
1.184
0.476
0.579
0.080
-0.595
NA
1.014
1.335
NA
100
100
NA
100
100
100
100
100
100
[47]
[49] test
[50]
-1.649
-0.535
-0.161
-0.926
-0.115
-0.142
1.721
1.310
1.139
1.467
1.116
1.130
[51]
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
[52]
[54]
1
3
[56]
[58]
NA – Not Available; № – indicates the number of a set in Table 3.
Creation of general training sets
In the next step of the study, the creation of the general training sets was performed using
different criteria. They were created by combining single datasets and included the following:
1
. A general training set, which was created by the combination of data from all thirteen
datasets in Table 3 (GeneralTr);
2. A training set that was created by combining the datasets mentioned in Table 4, for
which acceptable QSAR models were created (AcceptTr);
3
. A training set that was created by combining the data from the datasets in Table 3 with
the IC values calculated from the TxB2 blood level determination (TxB2Tr);
5
0
4
. A training set that was created by combining the data from the datasets with IC values
5
0
calculated from the TxB2 blood level determination, derived from the accepted QSAR
models mentioned in Table 4 (AcceptTxB2Tr);
5
. A training set that was created by combining the data from the datasets in Table 3 with
the IC values calculated from the determination of the PGE2 levels produced by the
5
0
action of hCOX-1 in vitro (PGE2Tr).
ACS Paragon Plus Environment
1
8

Page 19 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
A prediction of IC values was made for the external test set (Test set #4 from Table 3) by
5
0
QSAR modelling after the training of GUSAR based on the combined training sets (Table 5).
The combined training sets did not overlap with the test set.
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Table 5. Characteristics of QSAR models created by combining datasets.
2
tr
2
2
test
№
1
Training set
N
R
Q
RMSEtr
R
RMSEtest
GeneralTr
348
132
175
89
0.883
0.942
0.932
0.944
0.594
0.683
0.774
0.796
0.698
0.186
0.270
0.116
0.121
1.137
2
AcceptTr
0.577
0.446
0.473
0.952
1.005
0.997
3
TxB2Tr
4
AcceptTxB2Tr
5
PGE2Tr
160
There was no acceptable QSAR model
6
Consensus of prediction results given by the
single models from Table 4
0.067
1.021
N – number of structures in a training set.
Table 5 shows that GUSAR could not create any acceptable QSAR model based on the
PGE2Tr training set. The best prediction results was derived from a QSAR model created using
the AcceptTr training set, which contained the datasets of the QSAR single models mentioned in
2
Table 4. The values of R
and RMSEtest calculated from the QSAR models created using
test
AcceptTr, TxB2Tr and AcceptTxB2Tr were higher than the values for any single QSAR model
in Table 4 or their simple average consensus (the last row in Table 5). All the compounds from
the test set were in the applicability domain of the QSAR models based on the AcceptTr,
TxB2Tr and AcceptTxB2Tr training sets and 85% of the compounds were for GeneralTr.
The results shown in Table 5 were derived from the combined datasets. This simple
combination of datasets leads to the incorporation of duplicates of structures with different IC50
values. According to the general rules for QSAR modelling, duplicates in training sets should be
9
,10
9
avoided , and manual checking for duplicates and the deletion of error data is recommended .
ACS Paragon Plus Environment
1
9

Journal of Chemical Information and Modeling
Page 20 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
If both experimental properties are highly similar, maintaining the records associated with the
structure in the arithmetic average of properties is proposed. If the values are significantly
9
different, the elimination of both records is recommended . This is easily applied when analysing
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
the data from a single laboratory, measured using the same experimental protocol. However,
when one attempts to use as much data as possible to create a general training set from all the
available studies, it is difficult to recognize which experimental value among dozens is correct,
taking into account the special features of the experimental assays and the inevitable errors in
publications and databases. When we analysed duplicates in the combined datasets, six
compounds with duplicates were found, all of which corresponded to reference compounds. The
IC values of two reference compounds for the studied datasets are presented in Table 6.
5
0
Considering Table 6, it seems that all these data cannot be combined in the same training
set. The differences are the result of the use of different methods or different conditions when the
same method was used. This is one reason why the results obtained for GeneralTr were worst in
comparison with the other combined training sets. If one does not aim at analyzing the
dependency of the models on a specific parameter and includes limitations based on the value of
this parameter (such as the substrate concentration) in the selection of compounds for the training
set, this table may be used to choose which data can be combined. No one would feel safe
combining data when the values for the reference compound differ by an order of magnitude and
taking the average is not appropriate in this case. According to the table, it seems that the data
from sets 4, 6 and 2 may be combined and used for the prediction of IC values because they
5
0
were experimentally determined by the same method under the same conditions. This is the
AcceptTxB2Tr training set, which yields one of the better results in Table 5. Combining the data
ACS Paragon Plus Environment
2
0

Page 21 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
from sets 7, 8, 9 and 10 seems to be also acceptable. This is the PGE2Tr training set in Table 5.
However, in this case, an acceptable QSAR model could not be created.
Table 6. IC values of reference compounds in the studied datasets.
5
0
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Ref
Celecoxib, nM
23470
Indometacin, nM Scaling ratio
Product
0.639 TxB2
Source
1
2
3
4
5
6
7
8
9
1
[54]
[55]
[56]
[57]
[58]
[59]
[51]
[48]
[62]
[49]*
blood
7000
2.143 TxB2
1.071 TxB2
5.769 TxB2
1.056 TxB2
5.769 TxB2
PGE2
blood
14000
500
250
blood
2600
blood
14200
blood
2600
blood
100**
100
hCOX-1
hCOX-1
hCOX-1
hCOX-1
U937 microsomes
U-937 cells
PGE2
100
PGE2
0
15000
50
100**
1
PGE2
300 PGE2
2.941
11 [50]
[46]
1
2
5100
9336
7000
2
164
100
Average
Median
*
- Validation set; ** IC50 value using literature data for the same assay; Scaling ratio – a value used for the
multiplication of IC50 values to scale appropriate data to the IC50 value of Celecoxib (15000 nM – IC50 value of
Celecoxib in the test set); see below in the text for detailed explanation.
Table 6 shows that the experimental IC values of Celecoxib show a spread from 50 to
5
0
2
3470 nM. The minimum IC value for Celecoxib was determined by the measurement of the
50
PGE2 level produced by the enzyme derived from U937 microsomes. As mentioned above, the
use of the average value of duplicates is one of the approaches used to avoid duplicates. We
retrained GUSAR using the average value of duplicates in the combined datasets from Table 5.
Moreover, general training sets were also created from all the data (excluding the data from the
publication used for the test set) for compounds tested as inhibitors of human COX-1 from
ChEMBL (version 17) with average (CHEMBL ) or median (CHEMBLmed) IC50 values for
av
ACS Paragon Plus Environment
2
1

Journal of Chemical Information and Modeling
Page 22 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
duplicates. The results of the training and the prediction of IC values for the validation set are
5
0
shown in Table 7.
Table 7. Characteristics of QSAR models created by combining human COX-1 datasets
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
using average values for reference compounds.
2
2
2
№
1
Training set
GeneralTr
AcceptTr
N
R
Q
RMSEtr
0.600
R
RMSEtest
1.038
tr
test
333
130
169
87
0.932 0.618
0.946 0.714
0.942 0.769
0.945 0.797
0.187
0.300
0.023
0.017
2
0.551
0.884
3
TxB2Tr
0.456
1.006
4
AcceptTxB2Tr
0.477
1.098
5
6
7
8
PGE2Tr
157
There was no acceptable QSAR model
CHEMBLav
1137
1137
0.939 0.538
0.939 0.542
0.653
0.656
0.112
0.122
0.067
0.901 (AD 97%)
0.895 (AD 97%)
1.021
CHEMBLmed
Consensus of prediction results given
by the single models from Table 3
N – number of structures in a training set. CHEMBLav – dataset from all the compounds in ChEMBL tested for
human COX-1 inhibition, with addition of average IC50 values for the compounds having double/multiple records in
ChEMBL; CHEMBLmed – dataset from all the compounds in ChEMBL tested for human COX-1 inhibition, with
median IC values for the compounds having double/multiple records in ChEMBL.
50
Table 7 shows that QSAR models based on the GeneralTr and AcceptTr training sets were
improved by the averaging of the data. There were no changes in the accuracy of the prediction
for QSAR models based on TxB2Tr and PGE2Tr. The accuracy of prediction for QSAR models
based on the AcceptTxB2Tr training set was decreased. Interestingly, the accuracy of prediction
based on the CHEMBL and CHEMBL training sets was better than that of the other models.
av
med
The CHEMBL training set includes average IC values for compounds having double/multiple
av
50
records in ChEMBL. The CHEMBLmed training set includes median IC50 values for the
compounds having double/multiple records in ChEMBL. The general model based on the results
for all the recorded compounds using the median data for multiply recorded molecules achieved
the highest accuracy. The use of median values in training sets 1-5 (Table 7) did not lead to a
ACS Paragon Plus Environment
2
2

Page 23 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
change in the prediction results. This may be explained by the small number of duplicates in
these sets and the small difference between the given average and median values.
We introduce another approach for the combination of data from different sets – scaling,
which is based on analysis of experimental data of reference compounds. If we consider the IC50
values for reference compounds (Table 6), we can see that they show considerable differences.
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Figure 1A shows the distribution of IC values in the used sets of compounds tested for the
50
inhibition of human COX-1. The tenth set in Figure 1 is the test set. The black squares are the
IC values of Celecoxib, which is a reference compound in most studies, including the test set.
5
0
The idea of scaling is that the IC values of compounds in the sets are recalculated so as the IC
50
5
0
of the reference compound becomes equal in all the sets (Figure 1B). The datasets 7, 8, and 9
were not scaled because they didn’t contain the reference compound Celecoxib.
Figure 1. Distribution of IC values for compounds tested for the inhibition of human COX-1.
5
0
Black squares – IC value of Celecoxib. (A) Distribution of IC values based on the original
5
0
50
data. (B) Distribution of IC values after the scaling procedure.
5
0
ACS Paragon Plus Environment
2
3

Journal of Chemical Information and Modeling
Page 24 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
The scaling was accomplished by the multiplication of the IC values by an appropriate
5
0
scaling ratio (see the column “Scaling ratio” in Table 6). The scaling procedure leads to a change
in the general distribution of IC values in a combined training set. An example of this change is
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
5
0
shown in Figure 2.
Figure 2. General distribution of IC values in the sets.
5
0
One can see that after the scaling procedure the number of IC values that are close to the
5
0
average value increased (Figure 2). At the same time, the distribution of IC values in the test
5
0
set (tenth set in Figure 1) was shifted towards the less active compounds. The scaling procedure
was applied to the combined training sets mentioned in Table 6. Subsequently, new QSAR
models were created and tested on the external validation set (Table 8).
2
Comparison of the R
and RMSE values in Table 7 and Table 8 shows that for some training
test
sets the accuracy of the QSAR models increased after the scaling procedure (GeneralTr,
ACS Paragon Plus Environment
2
4

Page 25 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
AcceptTr, PGE2Tr), but it decreased for other training sets (TxB2Tr, AcceptTxB2Tr). In most
cases, the accuracy of the models trained on scaled data is higher than accuracy of the prediction
results given by the consensus of single models from Table 3.
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Table 8. Characteristics of the QSAR models created using the combined training sets after
data scaling to the reference compound (Celecoxib, IC = 15000 nM).
50
2
2
2
2
№
1
Training set
N
R
Q
Q
RMSEtr
R
test
RMSEtest
0.890
Sim
tr
Y-scr
GeneralTr
333
130
169
87
0.680
0.943
0.913
0.928
0.776
0.609
0.714
0.671
0.723
0.509
-0.257 0.633
-0.316 0.546
0.273
0.308
-0.046
0.082
0.074
0.067
0.337
0.299
0.332
0.511
0.361
2
AcceptTr
0.878
3
TxB2Tr
0.033
0.048
0.028
0.448
0.482
0.685
1.081
4
AcceptTxB2Tr
PGE2Tr
1.300
5
157
1.017
6
Consensus of prediction
results given by the single
models from Table 3
1.021
2
2
N – number of structures in a training set; Q
– average Q value calculated based on
Y-scr
internal validation. Sim – average value of pair wise similarity (Tanimoto coefficient) between
structures from the appropriate training set calculated based on MNA descriptors.
2
Table 8 shows that Q
values for all QSAR models on scaling data were less than 0.05.
Y-scr
2
This is significant less in comparison with Q values calculated based on the original data of the
2
training sets and indicates robustness of the models. The Q
values were better than, or
Y-scr
2
comparable to, R t values calculated for the test set by the QSAR models based on individual-
tes
dataset in Table 4.
To precisely evaluate the developed models that were based on the different designs
(individual, average, scaling) we performed a bootstrap analysis to determine if there were
statistically significant differences between the models:
ACS Paragon Plus Environment
2
5

Journal of Chemical Information and Modeling
Page 26 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
-
-
Each model was used to make a prediction for the same test set (dataset #4 in Table 3).
For this comparison to be fair, only results for chemical structures falling in the
applicability domain of the predictive model were used.
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
The results of each model’s prediction were bootstrapped 10 000 times using the R-
package called “resample” (https://CRAN.R-project.org/package=resample). The RMSE
value was calculated in each bootstrap run. Thus, many instances of RMSE
(https://CRAN.R-project.org/package=MLmetrics) were generated for each model.
-
RMSE values were used to conduct ANOVA in conjunction with Tukey’s HSD (honest
significance difference) test.
The results of this procedure indicated that models built using training sets of different
design, differed significantly (Pvalue < 0.05) from each other in terms of their quality as assessed
by RMSE (Figure 3). The illustration was prepared using ggplot2 package (https://cran.r-
project.org/web/packages/ggplot2/index.html). QSAR models based on the scaling procedure
showed the most accurate results.
Figure 3. Comparison of RMSE values given by the bootstrap analysis based on the
prediction results for the external test set (dataset #4 in Table 3). Individual means RMSE of the
ACS Paragon Plus Environment
2
6

Page 27 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
single models from Table 3. Average means RMSE of QSAR models (1-4 models from Table 7).
Scaling means RMSE of QSAR models (1-5 models from Table 8).
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
These approaches to using heterogeneous data for QSAR modelling based on published
experimental results were reproduced on ligand interactions with ovine COX-1, human COX-2
and soybean LOX (from ChEMBL using the same limitations as above for the available data on
the inhibitors of human COX-1). Novel experimental data concerning new
cyclooxygenase/lipoxygenase inhibitors synthesized at the School of Pharmacy of Aristotle
University of Thessaloniki and biologically evaluated in vitro were used as an external validation
set (Table 9) for the created QSAR models.
In Vitro COXs and LOX Inhibition Studies
The influence of the nature and the position of substitution of new thiazolidinone derivatives on
the cyclooxygenase and lipoxygenase inhibitory activities were evaluated experimentally (Table
9).
Table 9. Experimental IC50 values for COX-1, COX-2 and LOX for the studied
thiazolidinone derivatives.
O
HN
N
S
R
N
S
R1
IC , µM
50
compd
R
R1
LOX
COX-1,
COX-2,
ACS Paragon Plus Environment
2
7

Journal of Chemical Information and Modeling
Page 28 of 61
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
(% of inhibition at 10 µM)
(% of inhibition at 10 µM)
1
2
3
4
5
6
7
8
9
H
H
H
2-NO2
89.1
35.5
42.0
50.1
71.0
35.5
50.0
34.7
36.3
28.2
35.5
47.9
29.5
-
>200
0.3
- (9.4%)
- (9.4%)
- (31.7%)
- (7.9%)
- (58.8%)
- (32.2%)
- (20.3%)
- (1.0%)
- (20.6%)
- (26.4%)
- (6.6%)
-
H
3-NO2
H
4-NO2
0.5
H
2-Cl
0.018
22.4
0.3
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
H
3-Cl
H
4-Cl
H
3-OH
5.6
H
4-OH
7.9
1
0
1
H
4-OCH3
3-OMe, 4-OH
3,5-OMe, 4-OH
H
200
1
H
39.8
200
12
H
1
3
6- NO2
100 (61%)
63 (86%)
1
1
1
4
5
6
6- NO2
6- NO2
6- NO2
3- NO2
4- NO2
2-Cl
81.3
63.2
72.4
>200
79
106
>200
86
100 (12.7%)
17
6- NO2
6- NO2
6- NO2
6- NO2
6-OMe
6-OMe
6-OMe
4-CH3
4-CH3
4-CH3
4-Cl
4-OMe
34.7
63.1
26.3
33.9
33.9
35.9
42.7
63.1
38.0
200 (32%)
60 (125%)
>200
106 (61%)
200 (53%)
85 (71%)
>200 (25%)
>200
1
1
8
9
3-OMe, 4-OH
3,5-OMe, 4-OH
4- NO2
20
74 (95%)
74
2
2
2
1
2
3
3-Cl
200.5
>200
- (26.4%)
- (57%)
3-OMe, 4-OH
H
24
4
- (6.7%)
210
2
2
5
6
4-NO2
190
4-Cl
47.9
31.3
-
- (58%)
67 (86%)
Nordihydroguaiaretic acid
Celecoxib
-
-
34.5
0.1
Creation of QSAR models for ligands interacting with ovine COX-1, human COX-2 and
soybean LOX and their validation by external test sets with novel
cyclooxygenase/lipoxygenase inhibitors
QSAR modelling was performed based on 5 training sets for ovine COX-1 (Table 10), 44
training sets for human COX-2 (Table 11) and 3 training sets for soybean LOX (Table 12). The
created QSAR models had not been used to determine which compounds from Table 9 should be
ACS Paragon Plus Environment
2
8

Page 29 of 61
Journal of Chemical Information and Modeling
1
2
3
4
5
6
7
8
9
synthesized and experimentally studied as new cyclooxygenase/lipoxygenase inhibitors. The data
from Table 9 were used for external validation of the created QSAR models. The borders of the
IC values in Tables 10-12 are given without a scaling correction.
5
0
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
6
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
0
Table 10. Datasets of compounds tested as inhibitors of ovine COX-1.
No Product
Source
Nall Nsel
IC50min
nM)
IC50max log10(IC50max - IC50min
(nM)
)
Ref.
(
1
2
3
PGF2alpha Cayman kit
PGF2alpha Cayman kit
20 20
28 28
49 27
3200
231200
366000
390000
1.9
2.9
1.9
[60]
[61]
[62]
411
NA
seminal vesicles
5000
4
5
NA
NA
seminal vesicles
22 22
25 22
4620
50
1150000
17000
2.4
2.5
[63]
[64]
hematin-
reconstituted COX-1
N – number of all compounds studied in a publication; N - number of compounds with IC
50
all
sel
values that were selected for QSAR modelling. NA – Not Available.
Table 11. Datasets of compounds tested as inhibitors of human COX-2.
No Product
Source
Nall
Nsel
IC50min (nM) IC50max (nM) log10(IC50m Ref.
-
IC50min
)
ax
1
2
3
4
5
6
7
8
9
1
1
1
1
1
1
PGE2
PGE2
PGE2
PGE2
PGE2
PGE2
PGE2
PGE2
PGE2
PGE-2
NA
hCOX-2
blood
32 24
90 70
34 34
27 23
25 25
21 21
25 21
29 26
21 21
30 22
25 22
23 20
32 25
32 24
24 21
53 53
5.92
0.3
300
190
120
100
3000
6
10000
5720
3.2
4.3
2.1
3.5
2.2
2.3
1.8
2.4
3.2
3
[65]
[66]
[54]
[55]
[66]
[67]
[68]
[69]
[70]
[71]
[64]
[72]
[73]
[73]
[74]
[75]
blood
34610
540000
19400
19400
200000
1600
blood
hCOX-2
hCOX-2
hCOX-2
hCOX-2
CHO cell
ECV-304 cells
COX-2
2
2900
0
1
2
3
4
5
10
10000
66000
1900
40
3.2
3
Arachidonic acid CHO cells
2
NA
COX-2
blood
20
2240
2
NA
380
0.44
60
11300
388
1.5
2.9
5.1
PGE2
COS cells
blood
16 NA
7048553
ACS Paragon Plus Environment
2
9