Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

Contents lists available at ScienceDirect

Chemometrics and Intelligent Laboratory Systems

journal homepage: www.elsevier.com/locate/chemometrics

CORAL: QSAR models of CB1 cannabinoid receptor inhibitors based on local

and global SMILES attributes with the index of ideality of correlation and
the correlation contradiction index
Parvin Kumar a, *, Ashwani Kumar b
a
Department of Chemistry, Kurukshetra University, Kurukshetra, Haryana, 136119, India
b
Department of Pharmaceutical Sciences, Guru Jambheshwar University of Science and Technology, Hisar, Haryana, 125001, India

A R T I C L E I N F O A B S T R A C T

Keywords: Obesity has acquired notable attention due to its high occurrence and link with grievous health problems such as
QSAR hypertension, diabetes and heart disease. It has been reported that the endocannabinoid system executes a pivotal
SMILES part in the management of food absorption, fat augmentation, and energy balance. In the present manuscript, we
CORAL
report a detailed QSAR analysis for 165 CB1 cannabinoid receptor inhibitors employing the Monte Carlo opti-
CB1 receptor inhibitors
IIC
mization process incorporated within the CORAL software. Eight splits are made from the whole dataset and
CCI sixteen QSAR models are developed from these splits employing two target function TF1 (without index of ideality
of correlation) and TF2 (with index of ideality of correlation). All the QSAR models developed with TF2 have
better predictive potential than the models developed with TF1. The model built for split 5 using TF2 is the leading
model due to the higher value of the determination coefficient of the validation set (R2Valid ¼ 0.8518). The index of
ideality of correlation (IIC) improves the statistical performance of CORAL-based QSAR-models and gives sta-
tistically robust predictive models of the investigated endpoint pIC50. In the present manuscript, a novel criterion
“Correlation Contradiction Index (CCI)” is also applied to know its predictive potential. The absolute value of CCI
for calibration set is less when QSAR models are developed employing IIC. The promoters of increase and decrease
endpoint pIC50 are identified and these are applied to design seven new compounds. All the newly designed
molecule were docked into in the active site of human cannabinoid receptor CB1 (PDB ID: 5tgz).

that the endocannabinoid system executes a pivotal role in the man-

1. Introduction
Obesity is a worldwide health problem. In 2016, about 74 million
boys and 50 million girls were overweight and obesity has greatly
increased in recent decades [1]. Obesity has acquired notable attention
due to its high occurrence and link with grievous health problems such as
hypertension, diabetes and heart disease [2–4]. Food and Drug Admin-
istration (FDA) has approved the only one medicine, Orlistat, to treat the
obesity in children and adolescents. Whereas in adults, the FDA has
recommended six drugs i.e. Orlistat, Phentermine, Bupropion, Lorcaserin
and Liraglutide for obesity treatment. The therapeutic potential of these
drugs has been limited due to their unwanted side effects and variable
efficacy [5,6]. Therefore, the novel target development for the
anti-obesity drug is still a challenge to the medicinal chemist [7,8].
Literature survey reveals that the endocannabinoid system gave a new
platform for anti-obesity drug development [9–13]. It has been reported
agement of food absorption, fat augmentation, and energy balance [14].
The hyperactivation of this endogenous signalling system appears to be
firmly associated with abdominal obesity and the progress of the meta-
bolic syndrome. The two cannabinoid receptors subtypes, CB1 and CB2,
are a member of the G-protein coupled seven-transmembrane-spanning
receptor family (GPCRs) [15,16]. The CB1 and CB2 receptor are
expressed in the central nervous system & peripheral tissues and immune
system, respectively. The CB1 receptor is presumed to be associated in
the management of perception, motor activity, sensation, idea, under-
standing through thought, the senses, intuition resulting from the process
of cognition and the suppression of transmitter release through its
coupling to ions channels [17,18]. The antagonism of CB1 receptor has
been discovered by the scientific community as an alluring target for
obesity therapy because the specific blockage of this receptor can induce
bodyweight reduction [14,15,17].
Quantitative structure-activity relationship (QSAR) is an important

* Corresponding author.
E-mail addresses: parvinchem@kuk.ac.in, parvinjangra@gmail.com (P. Kumar).

https://doi.org/10.1016/j.chemolab.2020.103982
Received 23 July 2019; Received in revised form 24 February 2020; Accepted 28 February 2020
Available online 5 March 2020
0169-7439/© 2020 Elsevier B.V. All rights reserved.
P. Kumar, A. Kumar Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

a target function. This method is applied by the different research group

List of abbreviation for a large number of various physicochemical, biochemical, ecological,
and medicinal endpoints [19,23–25]. In the last few years, the index of
CB1 Cannabinoid Receptor-1 ideality correlation (IIC) has been calculated to judge the predictive
CCC Concordance Correlation Coefficient potential of built QSAR model and it helps to predict the better model in
CORAL: CORrelations And Logic place of various statistical parameters such as Q2, Q2F1, Q2F2, Q2F3 etc
IIC Index of Ideality Correlation [19,28–31].
MAE Mean Absolute Error In continuation to our work on QSAR [32–34] and synthesis of
QSAR Quantitative Structure Activity Relationship pharmacological active heterocyclic compounds [35–42], we herein
SMILES Simplified Input-Line Entry System publicize the CORAL based QSAR-models for 165 CB1 cannabinoid re-
CCI Correlation Contradiction Index ceptor inhibitors. In this manuscript, two new criteria Index of ideality
TF Target fuction correlation (IIC) and Correlation Contradiction Index (CCI) is also
studied.

2. Materials and method

part of computer-assisted drug design (CADD) and it has been effectively
applied for designing of new lead molecules [19–22]. QSAR correlates
the endpoints with the structural features of the molecules. Especially,
the CORAL software (http://www.insilico.eu/coral) based on the Monte
Carlo algorithm has been extensively used for designing the QSAR
models [19,20,23–27]. In this software, the molecular attributes ob-
tained from simplified input-line entry system (SMILES) and molecular
graph have been applied to develop QSAR models. The molecular
descriptor in terms of correlation weight (CW) has been calculated by
CORAL software and the numerical value of CW gives a maximal value of
2.1. Dataset
A total of 165 CB1 cannabinoid receptor inhibitors were taken from
the database https://www.ebi.ac.uk/chembl/old/(CheMBL ID are given
in Supporting information) and same experimental condition was applied
to measure the IC50 of CB1 receptor [6,17,43,44]. All IC50 (nM) values for
CB1 cannabinoid receptor were converted into the corresponding pIC50
(-logIC50) and it is taken as a dependent factor for the building of QSAR
models. Eight splits were made from the whole dataset and each split was

Table 1
Percentage of the identity of splits 1–8 CB1 receptor inhibitors.
Split SET Split 1 (%) Split 2 (%) Split 3 (%) Split 4 (%) Split 5 (%) Split 6 (%) Split 7 (%) Split 8 (%)

Split 1 Total 100 0.0 0.0 0.0 25.5 24.8 24.2 25.5
Training 100 0.0 0.0 0.0 24.4 26.5 24.4 24.4
Invisible training 100 0.0 0.0 0.0 26.2 24.1 24.1 26.5
Calibration 100 0.0 0.0 0.0 26.5 24.1 23.8 26.5
Validation 100 0.0 0.0 0.0 24.7 24.7 24.7 24.4

Split 2 Total 100 0.0 0.0 24.8 25.5 25.5 24.2

Training 100 0.0 0.0 24.1 26.2 26.5 24.1
Invisible training 100 0.0 0.0 26.5 24.4 24.4 24.4
Calibration 100 0.0 0.0 24.7 24.7 24.4 24.7
Validation 100 0.0 0.0 24.1 26.5 26.5 23.8

Split 3 Total 100 0.0 24.2 25.5 25.5 24.8

Training 100 0.0 24.7 24.4 24.7 24.7
Invisible training 100 0.0 23.8 26.5 26.5 24.1
Calibration 100 0.0 24.1 26.5 26.2 24.1
Validation 100 0.0 24.4 24.4 24.4 26.5

Split 4 Total 100 25.5 24.2 24.8 25.5

Training 100 26.5 23.8 24.1 26.5
Invisible training 100 24.4 24.7 24.7 24.7
Calibration 100 24.4 24.4 26.5 24.4
Validation 100 26.5 24.1 24.1 26.2

Split 5 Total 100 0.0 0.0 0.0

Training 100 0.0 0.0 0.0
Invisible training 100 0.0 0.0 0.0
Calibration 100 0.0 0.0 0.0
Validation 100 0.0 0.0 0.0

Split 6 Total 100 0.0 0.0

Training 100 0.0 0.0
Invisible training 100 0.0 0.0
Calibration 100 0.0 0.0
Validation 100 0.0 0.0

Split 7 Total 100 0.0

Training 100 0.0
Invisible training 100 0.0
Calibration 100 0.0
Validation 100 0.0

Split 8 Total 100

Training 100
Invisible training 100
Calibration 100
Validation 100

2
P. Kumar, A. Kumar Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

Fig. 1. Graphical representation of splits distribution.

further divided into four sets i.e. training set, invisible training set,
calibration set and validation set. Each set has its individual account-
ability in QSAR formation. The reported method [23,45]was applied to
know the non-identical character of splits and it is confirmed by Table 1.
The graphical presentation of split distribution is given in Fig. 1. Each set
has a discrete assignment to execute [25,46]. The assignment of the
training set was to make the QSAR model by calculating the correlation
weights and the assignment of the invisible training set was to test the
fitness of the compounds which were not present in training set. The
work assigned to the calibration set was to identify the start of over-
training whereas the job of the validation set was to verify the models for
the compounds which were not present in all the three sets i.e. training
set, invisible training set and calibration set [47,48].
Identity (%)¼ [Nij/0.5(NiþNj)]x100
The correlation weights in term of numerical data which give better
statistics for calibration set is calculated from the training and invisible
training set is represented by the following equation:
SMILES
Endpoint ¼ C0 þ C1  DCW (T, Nepoch) (2)
The predictive potential of the model is verified by the external
validation set.
In the CORAL software 2019, three types of target functions i.e. (i) the
classic scheme; (ii) balance of correlation; (iii) Index of Ideality of Cor-
relation (IIC) are designed. IIC has been recommended as a novel
benchmark and it is applied to judge the predictive potential of devel-
oped QSAR models. In this research paper, we use two target functions
for Monte Carlo optimization: balance of correlation (TF1) and balance of
correlation with IIC (TF2)

where: TF1 ¼ Rtrainingþ RinvTraining  │Rtraining  RinvTraining│ Const (3)

TF2¼ TF1 þ IIC Const (4)

Nij is the number of substances which are distributed into the same set
for both i-th split and j-th split (set ¼ training set, invisible training The Rtraining and RinvTraining are correlation coefficients between
set, calibration set, validation set) observed and calculated values of an activity/endpoint/property for the
Ni is the number of substances which are distributed into the set for i- training and invisible training sets, respectively. The empirical constant
th split (Const) is usually fixed.
Nj is the number of substances which are distributed into the set for j- It has been reported that the errors were found in a predicted QSAR/
th split QSPR model without Index of Ideality of Correlation (IIC) [19,29,50,51].
The following mathematical equation is applied to calculate the IIC for
any set:
2.2. Descriptor


min MAEset; MAEset
The descriptors used in developing the QSAR models for 165 CB1 IIC ¼ Rset 
 (5)
max MAEset; MAEset
cannabinoid receptor inhibitors based on SMILES notation are calculated
as reported below: The Rset is the correlation coefficient value between experimental and
predicted values of an endpoint for any used set.

 X X X
DCW T; Nepoch ¼ CWðSK Þþ CWðSSK Þ þ CWðSSSK Þ þ CWðBONDÞ þ CWðNOSPÞþ
(1)
CWðHALOÞ þ CWðHARDÞ þ CWðPAIRÞ þ CWðCmax Þ þ CWðNmax Þ þ CWðOmax Þ þ CWðSmax Þ

Here, CW(Z) is the correlation weight for a SMILE descriptor/attri- MAE is mean absolute error which can be determined with the
butes. T and Nepoch stand for threshold and number of epoch respectively. following equation:
Threshold coefficient is applied to classify various molecular attributes
1 X
N
obtained from SMILES into two classes: (i) active class i.e. CW is used in
M AEset ¼ jΔk j Δk < 0;N is the number of Δk < 0 (6)
the model making and (ii) rare class i.e. CW is not used in the model N k¼1
making. The best statistical quality for the calibration set is given by the
Nepoch. The detailed description of global SMILES attributes is given in
Table 2 [49].

3
P. Kumar, A. Kumar Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

Table 2 Table 3
The detailed description of SMILES attributes. The mathematical equation of various statistical parameters.
S.No. SMILES Comments S. No. Equation
notation P
1 ðYobs  Ypred Þ2
1. Sk SMILES-atoms, i.e. one symbol or two symbols which R2 ¼ 1  P
ðYobs  YÞ2
cannot be examined separately P
2 ðYobs  Ypred Þ2
2. SSk A combination of two SMILES-atoms 2
QLOO ¼ 1 P
ðYobs  YÞ2
3. SSSk A combination of three SMILES-atoms 8 92
3 > P >
4. BOND The presence or absence of double (‘ ¼ ’), triple (‘#’) and < next =
i¼1 ðYobs  Y obs ÞðYpred  Y pred Þ
stereochemical (‘@’) bonds R2ext ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ffi
>
: next 2 2> ;
5. NOSP Presence or absence of nitrogen, oxygen, sulfur and i¼1 ðY obs  Y obs Þ ðY pred  Y pred Þ
phosphorus Pnext 2
4 ðY obs  Y pred Þ
6. HALO Presence or absence of halogens
2
QF1 ¼ 1  Pni¼1 2
i¼1 ðYobs  Y train Þ
ext

7. PAIR The simultaneous presence of two SMILES-atoms from the Pnext 2

5 ðY obs  Y pred Þ
BOND, NOSP and HALO 2
QF2 ¼ 1  Pi¼1
next 2
8. Cmax Total number of rings (the range 0 .. 9) i¼1 ðYobs  Y ext Þ

9. Nmax Total number of nitrogen atoms in the molecular structure 6 minðMAEset; MAEsetÞ
IIC ¼ Rset 
10. Omax Total number of oxygen atoms in the molecular structure maxðMAEset; MAEsetÞ
11. Smax Total number of sulfur atoms in the molecular structure 7 Pntest ðyi  b y i Þ2
i¼1
ntest RMSEP2
2
QF3 ¼ 1 ¼ 1
Pntest ðyj  b y TR Þ2 S2TR

1 X
N j¼1
nTR
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 
M AEset ¼ jΔk j Δk < 0; þ N is the number of Δk < 0: (7) 8 2
rmðtestÞ ¼ r 2 ð1  r 2  r 2o  Þ
N k¼1 P
9 CCIcal ¼ ΔRj cal ; ΔRj cal < 0
P ext
The quality of prediction (Δk) for one substance from a set can be 10 2 ni¼1 ðYobs  Y obs ÞðYpred  Y pred Þ
CCC ¼ Pnext 2 Pnext 2 2
estimated as the following: i¼1 ðYobs  Y obs Þ þ i¼1 ðYpred  Y pred Þ þ next ðYobs  Y pred Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
11
CR2p ¼ R ðR  Rr Þ
2 2
Δk ¼ observedk – calculatedk (8)
12 1 P 
MAE ¼  Yobs  Ypred 
n rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
13 1 X
RMSEP ¼  ðYobs  Ypred Þ2
n
2.3. Correlation Contradiction Index (CCI)
the validation of the generated QSAR models [53–61]. The mathematical
The Correlation Contradiction Index (CCI) is a new criterion and it equation of various statistical parameters is given in Table 3. The IIC was
works on the hypothesis “The smaller absolute value of CCI, the better also implemented to validate the predictive potential and to judge the
predictive potential of a model (i.e. statistical characteristics for the better model for a developed QSAR model [19,29,50,51,62]. Finally, the
validation set should be better if absolute value of CCI is small)” given by novel criterion, CCI was also used to validate the built QSAR models.
Toropova et al. [52]. The CCI is the sum of all negative contributions over
the calibration set and it is represented by the following mathematical
equation 2.5. Applicability domain
X
CCIcal ¼ ΔRj cal ; ΔRj cal < 0 (9) From the list of various standard statistical parameter, applicability
domain (AD) is one of the most significant parameters which is procured
ΔRj cal ¼ R2cal  R2j cal (10) from the training set data of a particular in silico model. The AD is
summarized in the groups of structural, biological, physicochemical
where R2cal ¼ determination coefficient for all objects for calibration set; space, data or information and mechanism of action for which the model
of the training set is made and it can be applied for the evaluation of the
R2j cal ¼ correlation coefficient calculated without the jth object.
reliability of a built QSAR model. AD can be applied to judge the pre-
The jth object is “supporter” if the ΔRj cal is positive and its aid to the dictability of a developed QSAR model for the molecules which are not
correlation coefficient of the calibration set. Vice versa, the jth object is considered in making of the model. AD is useful to recognize the inter-
“oppositionist” if the ΔRj cal is negative and it decreases the correlation polation (true predictions) or extrapolation (less reliable predictions) of
coefficient of the calibration set. the developed QSAR model. However, in the case of the Monte Carlo
method based QSAR by CORAL, the statistical defects of SMILES are the
2.4. Validation principles according to which AD is defined. The “statistical defect,” d(A)
for an attribute of SMILES is helpful to understand the robustness of the
The main target of the developed QSAR model is to build a robust model and it can be represented by the following mathematical equation:
model which is efficient to predict the properties of novel molecule in an
jPðAÞ  P’ðAÞj
objective, authentic and precise manner. The judgment of robustness and dðAÞ ¼ (11)
NðAÞ þ N’ðAÞ
reliability of a built QSAR model can be made on the basis of the
following approach: a) internal validation or cross-validation taking into Here, P(A) & P0 (A) are the probabilities of an attribute (A) in the
account the training set data, b) external validation taking into account training and calibration set respectively; N(A) and N0 (A) are the number
the test set data and c) data randomization or Y-scrambling. The Monte of times attribute (A) appears the above sets.
Carlo based QSAR models are successfully validated by applying this For a given molecule, the statistical defect (D) is the summation of the
approach. statistical defect, d(A), of all the features available in the SMILES
Various standard statistical parameters such as correlation coefficient notation
(r2), concordance correlation coefficient (CCC), cross-validated correla-
tion coefficient (q2), Q2F1, Q2F2, Q2F3, standard error of estimation (s), X
NA
D ¼ defectðSMILESÞ ¼ dðAÞ (12)
mean absolute error (MAE), Fischer ratio (F) and root-mean-square error k¼1
(RMSE), and novel metrics (Rm2 and MAE based metric) are applied for

4
 P. Kumar, A. Kumar
Table 4
The various statistical parameter of built QSAR models with their comparisons for eight splits using TF1 and TF2.
Split Target SET n R2 CCC IIC Q2 Q2F1 Q2F2 Q2 F 3 s MAE F F (0.05, 1, p Y-test CR2p Avg Rm2 ΔRm2 Equation
Function n-2)

1 TF1 Training 41 0.9325 0.9651 0.8340 0.9249 0.228 0.184 539 251.06 0.034 0.0245 0.9202 pIC50 ¼ 3.9770599 (0.0310344) þ
InvTrain 42 0.9449 0.9697 0.8245 0.9385 0.231 0.176 686 251.14 0.030 0.0199 0.9349 0.0983955 (0.0007412) *
Calib 42 0.7136 0.8392 0.7846 0.6734 0.6408 0.6403 0.6217 0.548 0.413 100 251.14 0.079 0.0335 0.6966 0.5993 0.1573 DCW(10,6)
Valid 40 0.8033 0.8962 0.8447 0.7782 0.392 0.319 155 250.98 0.064 0.0307 0.7878 0.7205 0.0067

TF2 Training 41 0.8421 0.9143 0.8740 0.8230 0.348 0.267 208 251.06 0.055 0.0169 0.8336 pIC50 ¼ 2.4315214 (0.0748214) þ
InvTrain 42 0.9218 0.9538 0.7685 0.9128 0.266 0.218 471 251.14 0.036 0.0284 0.9075 0.0858929 (0.0011069) *
Calib 42 0.8045 0.8799 0.8969 0.7843 0.7975 0.7972 0.7868 0.412 0.318 165 251.14 0.062 0.0256 0.7916 0.6574 0.1914 DCW(3,12)
Valid 40 0.8187 0.8936 0.5760 0.7966 0.378 0.300 172 250.98 0.060 0.0303 0.8034 0.7376 0.1532

Comments Correct Incorrect Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Incorrect

2 TF1 Training 42 0.9674 0.9834 0.8941 0.9643 0.164 0.112 1186 251.14 0.023 0.0167 0.9590 pIC50 ¼ 1.9308274 (0.0274645) þ
InvTrain 41 0.9663 0.9795 0.9805 0.9618 0.171 0.125 1117 251.06 0.024 0.0411 0.9455 0.1330787 (0.0005774) * DCW(1,3)
Calib 40 0.7784 0.8710 0.7372 0.7557 0.7748 0.7748 0.7915 0.407 0.344 134 250.98 0.068 0.0377 0.7593 0.6673 0.1980
Valid 42 0.7126 0.8420 0.8310 0.6794 0.497 0.364 99 251.14 0.079 0.0215 0.7018 0.6004 0.1020

TF2 Training 42 0.9316 0.9646 0.9652 0.9246 0.237 0.169 545 251.14 0.034 0.0170 0.9231 pIC50 ¼ 1.2023358 (0.0499522) þ
InvTrain 41 0.8907 0.9385 0.9029 0.8794 0.296 0.237 318 251.06 0.044 0.0145 0.8834 0.1078466 (0.0007547) *
Calib 40 0.8917 0.9406 0.9436 0.8779 0.8898 0.8898 0.8979 0.285 0.219 313 250.98 0 .045 0.0170 0.8832 0.8259 0.1036 DCW(2,13)
Valid 42 0.7368 0.8524 0.7964 0.7104 0.470 0.382 112 251.14 0.075 0.0233 0.7251 0.6283 0.1692

Comments Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct

3 TF1 Training 40 0.8603 0.9249 0.9275 0.8436 0.321 0.254 234 250.98 0.052 0.0236 0.8484 pIC50 ¼ 3.8816131 (0.0509205) þ
InvTrain 42 0.8778 0.9282 0.7006 0.8600 0.328 0.258 287 251.14 0.047 0.0354 0.8599 0.1307608 (0.0015534) * DCW(6,2)
Calib 42 0.8359 0.9059 0.7129 0.8166 0.8342 0.8342 0.8272 0.369 0.281 204 251.14 0.055 0.0195 0.8261 0.7233 0.1575
Valid 41 0.7284 0.8527 0.7843 0.6992 0.467 0.364 105 251.06 0.077 0.0190 0.7188 0.6217 0.0574
5

TF2 Training 40 0.8794 0.9359 0.8485 0.8655 0.298 0.229 277 250.98 0.048 0.0189 0.8699 pIC50 ¼ 2.8318831 (0.0568819) þ
InvTrain 42 0.8682 0.9225 0.7112 0.8504 0.337 0.277 264 251.14 0.049 0.0197 0.8583 0.1401427 (0.0014856) * DCW(4,6)
Calib 42 0.8188 0.9016 0.9035 0.7972 0.8132 0.8132 0.8053 0.392 0.296 181 251.14 0.059 0.0076 0.8150 0.7405 0.0790
Valid 41 0.7379 0.8571 0.7858 0.7084 0.470 0.353 110 251.06 0.075 0.0178 0.7289 0.6345 0.0058

Comments Correct Correct Correct Incorrect Incorrect Incorrect Incorrect Incorrect Incorrect Incorrect Incorrect Correct Incorrect correct correct

4 TF1 Training 42 0.7961 0.8865 0.8922 0.7665 0.413 0.326 156 251.14 0.063 0.0193 0.7864 pIC50 ¼ 3.6452627 (0.0785066) þ
InvTrain 40 0.9038 0.9364 0.5578 0.8940 0.298 0.218 357 250.98 0.042 0.0094 0.8991 0.0919756 (0.0016254) * DCW(8,3)

Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

Calib 41 0.6975 0.8351 0.7512 0.6603 0.6723 0.6719 0.6799 0.502 0.378 90 251.06 0.083 0.0157 0.6896 0.5833 0.0057
Valid 42 0.8004 0.8875 0.8895 0.7748 0.431 0.338 160 251.14 0.062 0.0238 0.7884 0.7166 0.0212

TF2 Training 42 0.9126 0.9543 0.7892 0.8994 0.270 0.208 418 251.14 0.038 0.016 0.9044 pIC50 ¼ 1.0874582 (0.0761792) þ
InvTrain 40 0.9273 0.9454 0.3995 0.9193 0.298 0.231 485 250.98 0.036 0.017 0.9187 0.1013066 (0.0011163) *
Calib 41 0.7401 0.8570 0.8600 0.7139 0.6888 0.6884 0.6960 0.490 0.378 111 251.06 0.075 0.022 0.7289 0.6347 0.1257 DCW(2,14)
Valid 42 0.8220 0.9035 0.6921 0.7971 0.394 0.310 185 251.14 0.058 0.026 0.8088 0.7456 0.0275

Comments Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Incorrect Correct Correct Incorrect

5 TF1 Training 41 0.8821 0.9374 0.8945 0.8688 0.287 0.209 292 251.06 0.046 0.0399 0.8620 pIC50 ¼ 1.6195651 (0.0652232) þ
InvTrain 42 0.8802 0.9365 0.7799 0.8692 0.287 0.233 294 251.14 0.046 0.0427 0.8586 0.1138487 (0.0011623) * DCW(9,2)
Calib 41 0.7841 0.8776 0.6582 0.7587 0.7726 0.7588 0.6793 0.462 0.342 142 251.06 0 .066 0.0155 0.7764 0.6944 0.0605
Valid 41 0.8311 0.9094 0.7039 0.8081 0.389 0.315 192 251.06 0.057 0.0223 0.8199 0.7567 0.0967

TF2 Training 41 0.8767 0.9343 0.8918 0.8638 0.294 0.226 277 251.06 0.047 0.0209 0.8662 pIC50 ¼ 1.4915892 (0.0633723) þ
InvTrain 42 0.8720 0.9319 0.8522 0.8576 0.285 0.225 273 251.14 0.048 0.0123 0.8659 0.0755870 (0.0007347) * DCW(3,9)
Calib 41 0.8413 0.9010 0.9172 0.8237 0.8427 0.8332 0.7783 0.384 0.267 207 251.06 0.055 0.0285 0.8269 0.6813 0.1685
Valid 41 0.8518 0.9107 0.8864 0.8331 0.374 0.327 224 251.06 0 .053 0.0232 0.8401 0.7440 0.1440

Comments Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Incorrect Correct Incorrect Incorrect

6 TF1 Training 42 0.9584 0.9788 0.8088 0.9548 0.162 0.119 922 251.14 0.026 0.0357 0.9404
(continued on next page)
 P. Kumar, A. Kumar
Table 4 (continued )
Split Target SET n R2 CCC IIC Q2 Q2F1 Q2F2 Q2 F 3 s MAE F F (0.05, 1, p Y-test CR2p Avg Rm2 ΔRm2 Equation
Function n-2)

InvTrain 41 0.9331 0.9432 0.8925 0.9246 0.308 0.260 544 251.06 0.033 0.0192 0.9235 pIC50 ¼ 0.4873158 (0.0377398) þ
Calib 41 0.8389 0.9158 0.7818 0.8226 0.8474 0.8338 0.7786 0.384 0.301 203 251.06 0.056 0.0292 0.8242 0.7688 0.0208 0.1521567 (0.0006791) *
Valid 41 0.6395 0.7901 0.6863 0.5629 0.654 0.429 69 251.06 0.095 0.0317 0.6234 0.5062 0.1784 DCW(4,10)

TF2 Training 42 0.8380 0.9118 0.8322 0.8193 0.320 0.249 207 251.14 0.055 0.0477 0.8138 pIC50 ¼ 2.7870194 (0.0641000) þ
InvTrain 41 0.8147 0.9024 0.7803 0.7933 0.367 0.261 172 251.06 0.060 0.0561 0.7862 0.1057928 (0.0013248) * DCW(3,6)
Calib 41 0.8634 0.9109 0.9284 0.8465 0.8538 0.8408 0.7879 0.376 0.316 246 251.06 0.050 0.0271 0.8497 0.7639 0.1329
Valid 41 0.7136 0.8374 0.7536 0.6831 0.509 0.350 97 251.06 0.080 0.0270 0.7000 0.5994 0.1533

Comments Correct Correct Correct Correct Correct Correct Correct Correct Incorrect Correct Correct Correct Correct Incorrect Incorrect

7 TF1 Training 41 0.9059 0.9506 0.8220 0.8950 0.289 0.242 375 251.06 0.040 0.0118 0.8999 pIC50 ¼ 0.2012223 (0.0726707) þ
InvTrain 41 0.9124 0.9486 0.8030 0.9005 0.286 0.214 406 251.06 0.039 0.0183 0.9032 0.0965379 (0.0008883) *
Calib 42 0.6716 0.8117 0.4452 0.6372 0.6750 0.6381 0.7419 0.478 0.335 82 251.14 0.087 0.0107 0.6663 0.5497 0.0829 DCW(10,3)
Valid 41 0.6639 0.8045 0.5928 0.6283 0.538 0.406 77 251.06 0.090 0.0187 0.6545 0.5414 0.0368

TF2 Training 41 0.8739 0.9327 0.8073 0.8587 0.334 0.276 270 251.06 0.048 0.0194 0.8642 pIC50 ¼ 0.4618361 (0.0840195) þ
InvTrain 41 0.8636 0.9201 0.8769 0.8442 0.363 0.263 247 251.06 0.050 0.0292 0.8489 0.0934012 (0.0010513) *
Calib 42 0.7416 0.8584 0.8607 0.7151 0.7487 0.7201 0.8004 0.421 0.323 115 251.14 0.074 0.0086 0.7373 0.6390 0.0339 DCW(4,11)
Valid 41 0.7223 0.8392 0.6313 0.6913 0.4824 0.3706 101 251.06 0.079 0.0189 0.7128 0.6145 0.0059

Comments Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct

8 TF1 Training 41 0.9286 0.9630 0.9177 0.9187 0.251 0.193 507 251.06 0.035 0.0447 0.9059 pIC50 ¼ 0.7891190 (0.0833608) þ
InvTrain 41 0.9285 0.9633 0.9322 0.9206 0.254 0.205 507 251.06 0.035 0.0252 0.9158 0.1757479 (0.0016300) * DCW(7,2)
Calib 41 0.7007 0.8319 0.6708 0.6714 0.6699 0.6416 0.7171 0.501 0.355 91 251.06 0.082 0.0056 0.6979 0.5864 0.0655
Valid 42 0.6556 0.8067 0.4541 0.6096 0.482 0.354 76 251.14 0.091 0.0267 0.6421 0.5302 0.0801

TF2 Training 41 0.8436 0.9151 0.7188 0.8224 0.372 0.289 210 251.06 0.055 0.0119 0.8376 pIC50 ¼ 1.5466503 (0.0956255) þ
InvTrain 41 0.9166 0.9388 0.633 0.9076 0.327 0.277 429 251.06 0.038 0.0379 0.8974 0.1310703 (0.0018611) * DCW(3,5)
6

Calib 41 0.7398 0.8588 0.8564 0.7130 0.7331 0.7102 0.7713 0.450 0.343 111 251.06 0.075 0.0461 0.7164 0.6366 0.0434
Valid 42 0.7437 0.8615 0.6794 0.7136 0.411 0.301 116 251.14 0.073 0.0307 0.7282 0.6415 0.0464

Comments Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Incorrect Correct Correct Correct

Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

P. Kumar, A. Kumar Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

Fig. 2. Graphical representation of experimental pIC50 versus calculated pIC50 for eight models developed by TF2.

Table 5
The statistical characteristics of models for CB1 receptor inhibitors for eight random splits.
Split TF1 TF2

IICcalib CCIcalib │CCIcalib│ R2valid IICcalib CCIcalib │CCIcalib│ R2valid

1 0.7846 0.1263 0.1263 0.8033 0.8969 0.1087 0.1087 0.8187

2 0.7372 0.1363 0.1363 0.7126 0.9436 0.0638 0.0638 0.7368
3 0.7129 0.0879 0.0879 0.7284 0.9035 0.0051 0.0051 0.7379
4 0.7512 0.1606 0.1606 0.8004 0.8600 0.1489 0.1489 0.8220
5 0.6582 0.1266 0.1266 0.8311 0.9172 0.0980 0.0980 0.8518
6 0.7818 0.0992 0.0992 0.6395 0.9284 0.0770 0.0770 0.7136
7 0.4452 0.2007 0.2007 0.6639 0.8607 0.1619 0.1619 0.7223
8 0.6708 0.1892 0.1892 0.6556 0.8564 0.1532 0.1532 0.7437

3. Results and discussion

In the present manuscript, cannabinoid receptor-1 (CB1) activity

Fig. 3. Graphical representation of CCI and determination coefficient R2valid for
TF1 and TF2.
In CORAL, a molecule is categorized outlier if
D>2D
where D is the average of D calculated for training, invisible training and
calibration set. Thus for the validation set, if a molecule obeys equation
(12), then it will fall in the applicability domain.
(pIC50) was used as a function of molecular attributes extracted from
Simplified Molecular Input Line Entry System (SMILES) based de-
scriptors. A set of 165 CB1 were modelled in CORAL 2019 software
(CORAL software, IRCCS, Milan, Italy, 2019; available at: http://www.
insilico.eu/CORAL/) using Monte-Carlo optimization technique. Two
target functions were pondered to develop QSAR models viz. with and
without considering the influence of IIC in the determination of the
binding affinity for CB1 receptor. The range of 1–10 and 1 to 50 was
applied to get the preferable threshold value and number of epochs for
searching of the best T* and N*, respectively. The dRweight, IICweight and
Dlimit were 0.1, 0.2 and 0.1 respectively. The value for optimization Dstart
step was 1.0*CW(SA). Here, CW(SA) is the correlation weight of struc-
tural attribute (SA) at the start. Table 4 contains the calculated statistical
parameters for the QSAR models of CB1 receptor inhibitors based on
SMILES descriptors for eight random splits. The threshold and number of
epoch were selected to obtain the best statistical parameter for the cali-
(13)
bration set. In case of TF1, without taking into account the influence of
IIC on activity, the best (T*, N*) for splits 1, 2, 3, 4, 5, 6, 7 and 8 were
found to be (10,6), (1,3), (6,2), (8,3), (9,2), (4,10), (10,3) and (7,2)
respectively. On the other hand, in case of TF2, the optimal (T*, N*) were
(3,12), (2,13), (4,6), (2,14), (3,9), (3,6), (4,11) and (3,5) for splits 1, 2, 3,
4, 5, 6, 7 and 8 respectively by considering the influence of IIC on the
activity (pIC50). The above result clearly shows that the preferable

7
P. Kumar, A. Kumar Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

Table 6
Promoters of increase/decrease for endpoint according to selected "depth of interpretation".
S. SAk CWs Probe CWs Probe CWs Probe N1a N2a N3a Defect Comments
No. 1 2 3 [SAk]b

Promoter of endpoint increase

1 1….….… 2.75038 5.38376 3.76441 41 42 41 0 Presence of one ring
2 3….….… 4.21795 7.14817 5.27288 41 42 41 0 Presence of three ring
3 C…(….… 0.11638 0.27255 0.14882 41 42 41 0 Combination of sp3 carbon with branching
4 Cl….…... 0.01125 0.65255 0.38352 41 42 40 0.0003 Presence of chlorine
5 c….….… 0.22126 0.04514 0.21497 41 42 41 0 Presence of sp2 carbon or aromatic carbon
6 c…2….… 1.34719 0.6823 0.23938 41 40 37 0.0013 Presence of at least two aromatic ring
7 c…c…(… 0.69266 0.18772 0.46362 41 42 41 0 Presence of ethene or two sp2 carbon with branching
8 Cl..(….… 0.11575 0.618 0.04731 38 42 38 0 Presence of chlorine with branching
9 þþþþN⋯B2 ¼ ¼ 0.60519 0.35746 0.19689 35 41 32 0.0011 Presence of nitrogen with double bond
10 þþþþO⋯B2 ¼ ¼ 0.28592 0.36239 0.43363 35 41 32 0.0011 Presence of oxygen with double bond
11 NOSP11000000 1.10578 0.44356 0.91182 35 39 31 0.0015 Presence of nitrogen with oxygen
12 n … (… c … 1.27759 0.36719 0.36023 35 35 36 0.0003 Combination of sp2 nitrogen, branching and aromatic carbon
or sp2 carbon
13 4 …. …. … 1.42465 1.17113 2.31044 34 37 38 0.0014 Presence of four ring
14 c … 1 …. … 0.51551 0.02919 0.46233 31 35 30 0.0004 Presence of at least one aromatic ring
Promoter of endpoint decrease
1 ¼ …(….… 0.20398 1.70618 0.38374 35 41 32 0.0011 Combination of double bond and branching
2 ¼ …O…(… 1.3814 1.42981 1.23495 35 41 32 0.0011 Combination of double bond, oxygen and branching
3 O….…… 0.29202 0.07163 0.04181 35 41 32 0.0011 Presence of oxygen
4 O… ¼ …(… 1.6689 0.7186 0.44827 35 41 32 0.0011 Combination of oxygen, double bond and branching
5 O… ¼ ….… 1.36672 1.19371 0.35245 35 41 32 0.0011 Presence of doubly bonded oxygen
6 N⋯C….… 0.05738 0.37398 0.76953 33 37 28 0.002 Combination of sp3 nitrogen with sp3 carbon
7 c…n…c… 0.59709 0.68426 0.9763 28 31 34 0.0024 sp2 nitrogen surrounded by two sp2 carbon
8 n…c….… 0.64473 0.28789 1.30243 28 31 34 0.0024 Combination of sp2 nitrogen and sp2 carbon
9 C⋯C...1… 0.59633 0.23619 0.50714 10 7 11 0.0012 Presence of sp3 carbon in aliphatic ring
10 C⋯N⋯C … 0.99104 1.97473 0.49124 8 5 6 0.0035 sp3 nitrogen surrounded by two sp3 carbon
11 Cl..(…C… 3.12525 0.06177 0.212 7 10 6 0.0019 Combination of Chlorine, branching and sp3 carbon
12 þþþþCL–S ¼ ¼ 2.39359 1.09542 2.67555 6 3 8 0.0035 Presence of chlorine with sulfur
¼
13 þþþþN⋯S ¼ ¼ 1.34043 0.53547 0.54582 6 3 9 0.0049 Presence of nitrogen with sulfur
¼
14 4…c…2… 0.20206 0.97293 0.92051 5 11 9 0.007 Presence of four ring with two aromatic ring
a
The N1, N2 and N3 are the number of SMILES which contain given attribute (SAk) in training, invisible training and in calibration sets, respectively.
b
Defect [SAk] is the difference of probabilities of SAk in training and calibration sets, divided by sum of total numbers of the SAk in the training and calibration sets; If
attribute SAk absent in training set then defect ¼ 1 (maximum).

TF2 ¼ pIC50 ¼ 1.4915892 (0.0633723) þ 0.0755870 (0.0007347) *

threshold and number of epochs for all studied splits are not identical.
Total 16 QSAR models were developed using the balance of correlation
method with IIC (8 for TF2) and without IIC (8 for TF1). The IIC was also
applied to solve the task of a better model. The best QSAR models with
other statistical parameters are presented in Table 4 and in supporting
information.
The results presented in Table 4 clearly shows that all models are
statistically reliable and satisfied the criterion described by Tropsha et al.
[55] and Ojha et al. [63] for a predictive QSAR model. The statistical
parameters R2, CCC, IIC, Q2 (cross-validated correlation coefficient), s,
MAE, F, CR2p, Y-test, Q2F1, Q2F2, Q2F3, R2m (avg) and ΔR2m were
calculated to check the robustness and predictability of the generated
QSAR models. The Fisher-ratio (F), 0.05-quantile of Fisher’s distribution
F (0.05,1,n–2) and p-value (Fisher test’s significance level) has been re-
ported [64] in Table 4. These quantities were calculated using online
calculator (https://www.socscistatistics.com). The statistical parameters
R2 ¼ 0.8518, CCC ¼ 0.9107 and Q2 ¼ 0.8331 of the validation set of split
5 were best, therefore the built QSAR model of split 5 was selected as
preferred model. The values of various statistical parameters of split 5
with target function TF1 and TF2 are given below
TF1 ¼ pIC50 ¼ 1.6195651 (0.0652232) þ 0.1138487 (0.0011623) *
DCW(9,2)
n ¼ 41, R2 ¼ 0.8821, CCC ¼ 0.9374, IIC ¼ 0.8945, Q2 ¼ 0.8688
(Training set); n ¼ 42, R2 ¼ 0.8802, CCC ¼ 0.9365, IIC ¼ 0.7799, Q2 ¼
0.8692 (Invisible Training set); n ¼ 41, R2 ¼ 0.7841, CCC ¼ 0.8776, IIC
¼ 0.6582; Q2 ¼ 0.7587(Calibration set); n ¼ 41, R2 ¼ 0.8311; CCC ¼
0.9094; IIC ¼ 0.7039; Q2 ¼ 0.8081 (Validation set).
DCW(3,9)
n ¼ 41, R2 ¼ 0.8767, CCC ¼ 0.9343, IIC ¼ 0.8918, Q2 ¼ 0.8638
(Training set); n ¼ 42, R2 ¼ 0.8720, CCC ¼ 0.9319, IIC ¼ 0.8522, Q2 ¼
0.8576 (Invisible Training set); n ¼ 41, R2 ¼ 0.8413, CCC ¼ 0.9010, IIC
¼ 0.9172; Q2 ¼ 0.8237 (Calibration set); n ¼ 41, R2 ¼ 0.8518; CCC ¼
0.9107; IIC ¼ 0.8864; Q2 ¼ 0.8331 (Validation set).
The graphical representation of experimental pIC50 versus calculated
pIC50 for eight models developed by TF2 is given in Fig. 2. The QSAR
models for all splits premeditated by TF2 with the use of IIC are char-
acterized by good statistical attributes for the calibration and validation
set. These statistical experiments confirm that the predictive potential of
the models generated by the TF2 is better than the predictive potential of
the model build by TF1. This target function has been approved by the
outcome of the results of the various statistical parameter, as shown in
Table 4. The value of all statistical parameters calculated by the use of IIC
represented models correctly for splits 2 and 7. For split 1, all the sta-
tistical criteria except CCC and Δ R2m showed the model correctly. For
split 3, the statistical parameters R2, CCC, IIC, Y-test, avg R2m (test) and
Δ R2m showed better model correctly. For split 4, all statistical param-
eters except Δ R2m and Y-test showed the model correctly. For split 5, all
the statistical criteria except Y-test, avg R2m (test) and Δ R2m showed the
model correctly. For split 6, all the statistical criteria except MAE, avg
R2m (test) and Δ R2m showed the model correctly. For split 8, all the
statistical criteria except Y-test showed the model correctly. The result of
metric R2m and CCC showed that the developed QSAR models have good
predictive potential. The robustness of the developed QSAR models was
also established by the Y-randomization test (see supplementary infor-
mation). For the Y-randomization test, the value of CR2p was more than 0.5

8
 P. Kumar, A. Kumar
Table 7
Newly Designed molecules from CB018 and Lipinski’s rule of five.
S. Promoter of endpoint increase ID of designed SMILES Notation Structure calculated Lipinski’s rule of five
No. Compound pIC50
Molecular Number of Number of Virtual Lipinski
Weight Hydrogen Hydrogen logP Rule of
Bond acceptors Bond donors Five

CB018 (Lead Cc1c (nn (c2ccc (Cl)cc2Cl)c1c3ccc 6.9737 464.7721 3 1 6.396 75%
compound) (Cl)cc3)C (¼O)NC(¼O)C(C) (C)C

Combination of sp3 carbon with NCB01 CC(C)c1c (nn (c2ccc (Cl)cc2Cl) 7.1071 492.8253 3 1 7.035 75%
branching c1c3ccc (Cl)cc3)C (¼O)NC(¼O)
C(C) (C)C

Presence of chlorine NCB02 Cc1c (nn (c2ccc (Cl)cc2Cl)c1c3c 7.0819 499.2172 3 1 7.075 75%
(Cl)cc (Cl)cc3)C (¼O)NC(¼O)
C(C) (C)C
9

Combination of chlorine and sp3 NCB03 CC(C)c1c (nn (c2ccc (Cl)cc2Cl) 7.2153 527.2703 3 1 7.698 50%
carbon with branching c1c3c (Cl)cc (Cl)cc3)C (¼O)
NC(¼O)C(C) (C)C

Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

5. Aromatic carbon and Presence of at NCB04 C (c1ccccc1)c1c (nn (c2ccc (Cl) 8.6462 540.8680 3 1 7.851 50%
least two aromatic ring cc2Cl)c1c3ccc (Cl)cc3)C (¼O)
NC(¼O)C(C) (C)C

6. Aromatic ring and sp3 carbon with NCB05 CC(C)Cc1c (nn (c2ccc (Cl)cc2Cl) 7.9140 556.9105 3 1 8.318 50%
branching c1c3ccc (Cl)c4ccccc34)C (¼O)
NC(¼O)C(C) (C)C

7. Nitrogen with double bond NCB06 C (\N–

–C\c1ccccc1)c1c (nn (c2ccc 8.1522 567.8934 4 1 7.386 50%
(Cl)cc2Cl)c1c3ccc (Cl)cc3)C (¼O)
NC(¼O)C(C) (C)C
(continued on next page)
 P. Kumar, A. Kumar
Table 7 (continued )
S. Promoter of endpoint increase ID of designed SMILES Notation Structure calculated Lipinski’s rule of five
No. Compound pIC50
Molecular Number of Number of Virtual Lipinski
Weight Hydrogen Hydrogen logP Rule of
Bond acceptors Bond donors Five

8 Combination of Oxygen with NCB07 C (\N–

–C\c1ccccc1)c1c (nn (c2ccc 9.5048 657.9729 5 1 7.531 50%
double bond, aromatic system, (Cl)cc2Cl)c1c3ccc (Cl)cc3)C (¼O)
Nitrogen with double bond and NC(¼O)C(C) (C)C (¼O)c1ccccc1
Presence of at least two aromatic
ring

9 FDA Approved Drug Otenabant [66] CCNC1(CCN(CC1)c1c2nc (n 9.0297 510.4183 2 7 5.3793 50%
(c2ncn1)c1ccc (Cl)cc1) (Expt
c1ccccc1Cl)C (¼O)N 9.1549)
10

10 Surinabant O¼C(NN1CCCCC1)c1nn (c2ccc 7.6639 522.26492 1 4 6.6305 50%

[67] (Cl)cc2Cl)c (c2ccc (Br)cc2)c1CC (Expt
7.6576)

11 Ibipinament c1cc (Cl)ccc1C1 ¼ NN(C 6.7184 501.42808 1 6 5.6489 50%

(SLV319) [68] (¼NCS(¼O) (¼O)c2ccc (Cl)cc2) (Expt

Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

NC)CC1c1ccccc1 7.6576)
P. Kumar, A. Kumar
Fig. 4. Binding Interactions of newly designed compounds NCB01-NCB07 with lead molecule CB018.
[32]. The IIC was considered as a final parameter to validate the devel-
oped QSAR model and all models were correctly predicted by IIC. The
literature survey also reveals that “the more IIC, the higher statistical
quality of a model for external validation set” [19,29,50,51]. Hence it can
be generalized that "the IIC is a consistent criterion of predictive potential
of a model".
The mathematical equation (9) is applied to calculate CCI and it can
be employed as a criterion of predictive potential of a developed QSAR
11
Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982
model. The results of CCI are depicted in Table 5 and it shows that
smaller the absolute value of CCI, better is the model (more R2valid value).
The absolute value of CCI for the QSAR model developed with TF2 is less
than the QSAR model developed with TF1 (Table 5, Fig. 3). Literature
survey reveals that CCI is more correlated with R2 of the validation set in
comparison to IIC and in the present manuscript, this outcome is also
found true [52]. The absolute value of the correlation of R2valid with
CCIcalib (R ¼ 0.2013) is more than the correlation with IICcalib (R ¼
P. Kumar, A. Kumar
0.0809). Hence we may conclude that CCI has the predictive potential for
the models developed with IIC and it can be used as a statistical criterion
to predict a better model. However, this statement has been tested only
for the present eight split and this assumption should be verified in future
with different data set.
An important criterion of OECD principles is the applicability domain
(AD). Compounds falling outside the applicability domain are defined as
outliers. In case of TF1, without taking into account the influence of IIC
on activity (pIC50), the number of outliers for splits 1, 2, 3, 4, 5, 6, 7 and 8
are 0, 10, 0, 0, 0, 21, 0 and 14 respectively. On the other hand, In case of
TF2, the number of outlier are 0, 8, 17, 11, 22, 26, 16 and 19 for splits 1,
2, 3, 4, 5, 6, 7 and 8 respectively by considering the influence of IIC on the
pIC50 (see supplementary information).
3.1. Mechanistic interpretation, designing of the new molecules and
prediction of market drugs
A mechanistic interpretation is another significant criterion given by
OECD, which means that molecular features are accountable for the in-
crease and decrease of an endpoint can be extracted from such models. In
Monte Carlo optimization method, the optimal descriptor, correlation
weight values (CW), can be correlated with appropriate structural attri-
butes (SAk) which aid in the detection of the promoter of increase and
decrease of the endpoint.
The following three rules are generally applied to categories the
structural attributes achieved from three independent Monte Carlo
optimization runs as promoter of increase or decrease for endpoint: (i)
the structure attribute is promotor of increase if CW(SAk) is positive in all
runs (ii) the structure attribute is promotor of decrease if CW(SAk) is
negative in all runs (iii) if CW(SAk) is both positive and negative in all
runs, then that structure attribute is undefined. The list of all structural
attributes with their correlation weights for three runs of the developed
QSAR model for split 5 by using TF2 is depicted in Table 6. Seven new
cannabinoid CB1 receptor inhibitors (NCB01-NCB07) were designed
from the lead compound CB018. All the newly designed molecules with
their chemical structure, SMILES notation, predicted pIC50 and promotor
of endpoint increase is given in Table 7. The range of predicted pIC50 for
all the newly designed compounds is 7.0819–9.5048. The predicted
pIC50 of the most active designed compound is 9.5048. The Lipinski,s
rule of 5 was also applied to the designed molecules. The compound
CB01 and CB02 follows 75% Lipinski,s rule of 5, whereas for other
compounds follows 50% Lipinski,s rule [65] of 5. In order to check the
predictive potential of the model, we determined the endpoints of a three
market drugs such as Otenabant [66], Surinabant [67] and Ibipinament
(SLV319) [68]. The endpoint calculated for these drugs are very close to
the experimental endpoint (Table 7; entry 9–11)
3.2. Molecular docking studies
To further validate the design strategy, docking studies of base
compound CB018 and designed compounds was carried out with the help
of Autodock Vina program [69] and results were visualized with Dis-
covery Studio Visualizer. The 3D structure of molecules was prepared in
MarvinSketch and CB1 PDB (pdb id: 5tgz) was retrieved from Protein
Database (https://www.rcsb.org/structure/5TGZ) [70]. The compounds
were docked in the active site of human cannabinoid receptor CB1 (PDB
ID: 5tgz) and all the compounds emulated the extended anchoring
conformation of co-crystalized ligand AM6538. These compounds were
affixed pretty low in the active site of receptor nearly above Trp356 and
capped with Met103 (Fig. 4). The binding conformations of compounds
were stabilized by hydrophobic interactions mainly. Dichlorophenyl ring
was detected in the slender side pocket and made T shaped pi-pi in-
teractions with Phe170 which is a very important interaction for the
activity of compounds [70]. Addition of positive structural attributes
resulted in the better fitting of compounds by making suitable in-
teractions. In compound NCB01, isopropyl group led to more pi alkyl
12
Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982
interactions with active site residues as compared to the methyl group in
CB0018. Additional chloro group in compound NCB02 exhibited hy-
drophobic interactions with Leu359, Phe379 and Ser383. Presence of
both of these attributes in compound NCB03 headed towards more
number of interactions. Increase in the number of aromatic rings as in
compound NCB04 resulted in an increase in the number of pi-pi in-
teractions (pi-pi stacked and T-shaped) which caused the higher activity
of this molecule. However, changing the position of the aromatic ring
(NBC05) led to a decrease in the activity but it was still higher than the
base compound due to more number of interactions. Addition of nitrogen
with a double bond (molecule NCB06) also caused a better fit of the
molecule as compared to CB018 but it was less active than compound
NCB04. Chloro group of this compound also exhibited one halogen bond
with Gly166. Combination of oxygen with a double bond, aromatic
system, nitrogen with double bond and presence of at least two aromatic
ring formed compound NBC07 which came out to be most active due to
these attributes. Pyrazole ring of this compound displayed pi-sulfur
interaction with Met384 while iminophenyl ring made pi-donor
hydrogen bond with NH of Met103. These additional engagements
might be the cause of the higher activity of this molecule. Except for
compound NCB02 and NCB05, all other molecules also exhibited
T-shaped pi-pi interaction with Phe268 (Fig. 3). The residues involved in
binding interactions with these designed molecules are also involved in
binding with known antagonists namely rimonabant, otenabant, and
taranabant [70].
4. Conclusion
Here we report a detailed QSAR analysis for 165 CB1 receptor in-
hibitors employing the Monte Carlo optimization process incorporated
within the CORAL software. This method is applied to develop 16 models
for eight different random splits using two target function TF1 (without
IIC) and TF2 (with IIC). All the models developed with TF2 have better
predictive potential than the models developed with TF1. The model built
for split 5 using TF2 is the leading model due to the higher value of the
determination coefficient of the validation set (R2Valid ¼ 0.8518). The IIC
improves the statistical performance of CORAL-based QSAR-models and
gives statistically robust predictive models of the investigated endpoint
pIC50. In the present manuscript, a novel criterion Correlation Contra-
diction Index is also applied to know its predictive potential. The abso-
lute value of CCI for calibration set is less when QSAR models are
developed employing IIC. The promoters of increase and decrease
endpoint pIC50 are identified and these are applied to design seven new
compounds (NCB01-NCB07) and these are docked in the active site of
human cannabinoid receptor CB1 (PDB ID: 5tgz). The binding confor-
mations of compounds (NCB01-NCB07) were stabilized by hydrophobic
interactions mainly. Dichlorophenyl ring was detected in the slender side
pocket and made T shaped pi-pi interactions with Phe170 which is a very
important interaction for the activity of compounds. The residues
involved in binding interactions with these designed molecules are also
involved in binding with known antagonists namely rimonabant, ote-
nabant, and taranabant [70].
Declaration of competing interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
CRediT authorship contribution statement
Parvin Kumar: Conceptualization, Data curation, Methodology,
Software, Writing - original draft, Writing - review & editing. Ashwani
Kumar: Conceptualization, Data curation, Formal analysis, Investiga-
tion, Methodology, Software, Validation, Visualization, Writing - review
P. Kumar, A. Kumar Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982

& editing. pyrophosphate synthase inhibitors based on Monte Carlo method, Drug Res. 69
(2019) 159–167.
[21] A. Kumar, E. Rathi, S.G. Kini, Identification of potential tumour-associated carbonic
Acknowledgements anhydrase isozyme IX inhibitors: atom-based 3D-QSAR modelling, pharmacophore-
based virtual screening and molecular docking studies, J. Biomol. Struct. Dyn.
The authors are thankful to Dr Andrey A. Toropov and Dr Alla P. (2019), https://doi.org/10.1080/07391102.2019.1626285.
[22] M. Marzo, G.J. Lavado, F. Como, A.P. Toropova, A.A. Toropov, D. Baderna,
Toropova for providing CORAL software. The authors are also thankful to C. Cappelli, A. Lombardo, C. Toma, M. Blazquez, E. Benfenati, QSAR models for
their respective universities for providing the infrastructure. biocides: the example of the prediction of Daphnia magna acute toxicity, SAR QSAR
Environ. Res. 31 (2020) 227–243.
[23] A.A. Toropov, A.P. Toropova, The Monte Carlo Method as a tool to build up
Appendix A. Supplementary data predictive QSPR/QSAR, Curr. Comput. Aided Drug Des. (2019), https://doi.org/
10.2174/1573409915666190328123112.
Supplementary data to this article can be found online at https:// [24] A.P. Toropova, A.A. Toropov, CORAL: QSAR models for carcinogenicity of organic
compounds for male and female rats, Comput. Biol. Chem. 72 (2018) 26–32.
doi.org/10.1016/j.chemolab.2020.103982. [25] P. Kumar, A. Kumar, Monte Carlo method based QSAR studies of mer kinase
inhibitors in compliance with OECD principles, Drug Res. 68 (2018) 189–195.
References [26] A.P. Toropova, A.A. Toropov, E. Carnesecchi, E. Benfenati, J.L. Dorne, The index of
ideality of correlation: models for flammability of binary liquid mixtures, Chem.
Pap. 74 (2020) 601–609.
[1] F. Rossi, F. Punzo, G. Umano, M. Argenziano, E. Miraglia Del Giudice, Role of
[27] S. Jain, S.A. Amin, N. Adhikari, T. Jha, S. Gayen, Good and bad molecular
cannabinoids in obesity, Int. J. Mol. Sci. 19 (2018) 2690.
fingerprints for human rhinovirus 3C protease inhibition: identification, validation,
[2] A. Tremblay, E. Doucet, Obesity: a disease or a biological adaptation? Obes. Rev. 1
and application in designing of new inhibitors through Monte Carlo-based QSAR
(2000) 27–35.
study, J. Biomol. Struct. Dyn. 38 (2020) 66–77.
[3] G.A. Bray, Obesity: the disease, J. Med. Chem. 49 (2006) 4001–4007.
[28] A.P. Toropova, A.A. Toropov, A.M. Veselinovic, J.B. Veselinovic, D. Leszczynska,
[4] P.N. Patel, R. Pathak, Rimonabant: a novel selective cannabinoid-1 receptor
J. Leszczynski, Semi-correlations combined with the index of ideality of correlation:
antagonist for treatment of obesity, Am. J. Health Syst. Pharm. 64 (2007) 481–489.
a tool to build up model of mutagenic potential, Mol. Cell. Biochem. 452 (2019)
[5] D.R. Hou, S. Alam, T.C. Kuan, M. Ramanathan, T.P. Lin, M.S. Hung, 1,2,3-Triazole
133–140.
derivatives as new cannabinoid CB1 receptor antagonists, Bioorg. Med. Chem. Lett
[29] A.P. Toropova, A.A. Toropov, The index of ideality of correlation: improvement of
19 (2009) 1022–1025.
models for toxicity to algae, Nat. Prod. Res. 33 (15) (2019) 2200–2207, https://
[6] H.J. Seo, M.J. Kim, S.H. Lee, S.-H. Lee, M.E. Jung, M.-S. Kim, K. Ahn, J. Kim, J. Lee,
doi.org/10.1080/14786419.2018.1493591.
Synthesis and structure–activity relationship of 1,2,4-triazole-containing
[30] A.A. Toropov, A.P. Toropova, Predicting cytotoxicity of 2-phenylindole derivatives
diarylpyrazolyl carboxamide as CB1 cannabinoid receptor–ligand, Bioorg. Med.
against breast cancer cells using index of ideality of correlation, Anticancer Res. 38
Chem. 18 (2010) 1149–1162.
(2018) 6189–6194.
[7] K.A. Horton, A.V. Goonawardena, J. Sesay, A.C. Howlett, R.E. Hampson, Systemic
[31] S. Ahmadi, Mathematical modeling of cytotoxicity of metal oxide nanoparticles
blockade of the CB1 receptor augments hippocampal gene expression involved in
using the index of ideality correlation criteria, Chemosphere 242 (2020) 125192.
synaptic plasticity but perturbs hippocampus-dependent learning task, Cannabis
[32] P. Kumar, A. Kumar, J. Sindhu, In silico design of diacylglycerol acyltransferase-1
Cannabinoid Res. 4 (2019) 33–41.
(DGAT1) inhibitors based on SMILES descriptors using Monte-Carlo method,, SAR
[8] S.D. Banister, K. Krishna Kumar, V. Kumar, B.K. Kobilka, S.V. Malhotra, Selective
QSAR Environ. Res. 30 (2019) 525–541, https://doi.org/10.1080/
modulation of the cannabinoid type 1 (CB1) receptor as an emerging platform for
1062936X.2019.1629998.
the treatment of neuropathic pain, Medchemcomm 10 (2019) 647–659.
[33] Manisha, S. Chauhan, P. Kumar, A. Kumar, Development of prediction model for
[9] K. Yoshida, Y. Kita, S.M. Tokuoka, F. Hamano, M. Yamazaki, K. Sakimura, M. Kano,
fructose- 1,6- bisphosphatase inhibitors using the Monte Carlo method, SAR QSAR
T. Shimizu, Monoacylglycerol lipase deficiency affects diet-induced obesity, fat
Environ. Res. 30 (2019) 145–159.
absorption, and feeding behavior in CB1 cannabinoid receptor-deficient mice,
[34] A. Kumar, S. Singh, S. Jain, P. Kumar, Synthesis, antimicrobial evaluation, QSAR
Faseb. J. 33 (2019) 2484–2497.
and in Silico ADMET studies of decanoic acid derivatives, Acta Pol. Pharm. 68
[10] X. Xu, S. Jiang, E. Xu, X. Wu, R. Zhao, Inhibition of CB1 receptor ameliorates spatial
(2011) 191–204.
learning and memory impairment in mice with traumatic brain injury, Neurosci.
[35] P. Kumar, M. Duhan, K. Kadyan, J.K. Bhardwaj, P. Saraf, M. Mittal, Multicomponent
Lett. 696 (2019) 127–131.
synthesis of some molecular hybrid containing thiazole pyrazole as apoptosis
[11] M.C.G. Leite-Avalca, F.T. Staats, D. Verona, P. de Souza, M.C. Almeida, J.E. Silva-
inducer, Drug Res. 68 (2018) 72–79.
Santos, A.R. Zampronio, Cannabinoid CB1 receptor antagonist rimonabant
[36] P. Kumar, K. Kadyan, M. Duhan, J. Sindhu, V. Singh, B.S. Saharan, Design,
decreases levels of markers of organ dysfunction and alters vascular reactivity in
synthesis, conformational and molecular docking study of some novel acyl
aortic vessels in late sepsis in rats, Inflammation 42 (2019) 618–627.
hydrazone based molecular hybrids as antimalarial and antimicrobial agents, Chem.
[12] J.L.C. Lee, F.E. Amorim, L.F. Cassini, O.B. Amaral, Different temporal windows for
Cent. J. 11 (2017) 115.
CB1 receptor involvement in contextual fear memory destabilisation in the
[37] P. Kumar, M. Duhan, K. Kadyan, J. Sindhu, S. Kumar, H. Sharma, Synthesis of novel
amygdala and hippocampus, PloS One 14 (2019), e0205781.
inhibitors of α-amylase based on thiazolidine-4-one skeleton containing pyrazole
[13] H. Laurikainen, L. Tuominen, M. Tikka, H. Merisaari, R.L. Armio, E. Sormunen,
moiety and their configurational studies, MedChemComm 8 (2017) 1468–1476,
F. Borgan, M. Veronese, O. Howes, M. Haaparanta-Solin, O. Solin, J. Hietala,
https://doi.org/10.1039/C7MD00080D.
M. group, Sex difference in brain CB1 receptor availability in man, Neuroimage 184
[38] P. Kumar, R. Bhatia, R. Khanna, A. Dalal, D. Kumar, P. Surain, R.C. Kamboj,
(2019) 834–842.
Synthesis of some benzothiazoles by developing a new protocol using urea nitrate
[14] B.K. Srivastava, R. Soni, A. Joharapurkar, K.V. Sairam, J.Z. Patel, A. Goswami,
as a catalyst and their antimicrobial activities, J. Sulfur Chem. 38 (2017) 585–596.
S.A. Shedage, S.S. Kar, R.P. Salunke, S.B. Gugale, A. Dhawas, P. Kadam, B. Mishra,
[39] P. Kumar, R. Bhatia, D. Kumar, R.C. Kamboj, S. Kumar, R. Kamal, R. Kumar, An
N. Sadhwani, V.B. Unadkat, P. Mitra, M.R. Jain, P.R. Patel, Bioisosteric replacement
economic, simple and convenient synthesis of 2-aryl/heteroaryl/styryl/alkylben-
of dihydropyrazole of 4S-(-)-3-(4-chlorophenyl)-N-methyl-N’-[(4-chlorophenyl)-
zothiazoles using SiO 2–HNO 3, Res. Chem. Intermed. 41 (2015) 4283–4292.
sulfonyl]-4-phenyl-4,5-di hydro-1H-pyrazole-1-caboxamidine (SLV-319) a potent
[40] S. Kumar, S. Kumar, P. Kumar, Synthesis and antimicrobial activity of some (3-
CB1 receptor antagonist by imidazole and oxazole, Bioorg. Med. Chem. Lett 18
phenyl-5-(1-phenyl-3-aryl-1H-pyrazol-4-yl)-4, 5-dihydro-1H-pyrazol-1-yl)(pyridin-
(2008) 963–968.
4-yl) methanones: new derivatives of 1, 3, 5-trisubstituted pyrazolines, Med. Chem.
[15] C. Montero, N.E. Campillo, P. Goya, J.A. Paez, Homology models of the cannabinoid
Res. 22 (2013) 433–439.
CB1 and CB2 receptors. A docking analysis study, Eur. J. Med. Chem. 40 (2005)
[41] P. Kumar, A. Kumar, J. Makrandi, Synthesis and evaluation of bioactivity of
75–83.
thiazolo [3, 2-b]-[1, 2, 4]-triazoles and isomeric thiazolo [2, 3-c]-[1, 2, 4]-triazoles,
[16] C.-J. Qiao, H.I. Ali, K.H. Ahn, S. Kolluru, D.A. Kendall, D. Lu, Synthesis and
J. Heterocycl. Chem. 50 (2013) 1223–1229.
biological evaluation of indole-2-carboxamides bearing photoactivatable
[42] P. Kumar, A. Kuamr, L.J. Mohan, J. Makrandi, Heterocyclic systems containing
functionalities as novel allosteric modulators for the cannabinoid CB1 receptor, Eur.
bridgehead nitrogen atom: synthesis and evaluation of biological activity of
J. Med. Chem. 121 (2016) 517–529.
imidazo [2, 1-b]-1, 3, 4-thiadiazolo [2, 3-c]-s-triazoles, s-triazolo [3, 4-b]-1, 3, 4-
[17] J. Lee, H.J. Seo, S.H. Lee, J. Kim, M.E. Jung, S.-H. Lee, K.-S. Song, J. Lee, S.Y. Kang,
thiadiazolo [3, 2-b] imidazo [4, 5-b] quinoxaline and bis-(s-Triazolo [3, 4-b]-1, 3, 4-
M.J. Kim, M.-S. Kim, E.-J. Son, M. Lee, H.-K. Han, Discovery of 2-(4-((1H-1,2,4-
thiadiazolo [3, 2-b][imidazo [4, 5-b]-cyclohexane]-5a, 6a-diene), Bull. Kor. Chem.
triazol-1-yl)methyl)-5-(4-bromophenyl)-1-(2-chlorophenyl)-1H-pyrazol-3-yl)-5-
Soc. 31 (2010) 3304–3308.
tert-butyl-1,3,4-thiadiazole (GCC2680) as a potent, selective and orally efficacious
[43] K.-S. Song, M.J. Kim, H.J. Seo, S.-H. Lee, M.E. Jung, S.-U. Kim, J. Kim, J. Lee,
cannabinoid-1 receptor antagonist, Bioorg. Med. Chem. 18 (2010) 6377–6388.
Synthesis and structure–activity relationship of novel diarylpyrazole imide
[18] M.K. Sharma, P.R. Murumkar, A.M. Kanhed, R. Giridhar, M.R. Yadav, Prospective
analogues as CB1 cannabinoid receptor ligands, Bioorg. Med. Chem. 17 (2009)
therapeutic agents for obesity: molecular modification approaches of centrally and
3080–3092.
peripherally acting selective cannabinoid 1 receptor antagonists, Eur. J. Med.
[44] N. Khan, S.A. Halim, W. Khan, S.K. Zafar, Z. Ul-Haq, In-silico designing and
Chem. 79 (2014) 298–339.
characterization of binding modes of two novel inhibitors for CB1 receptor against
[19] P. Kumar, A. Kumar, J. Sindhu, Design and development of novel focal adhesion
obesity by classical 3D-QSAR approach, J. Mol. Graph. Model. 89 (2019) 199–214.
kinase (FAK) inhibitors using Monte Carlo method with index of ideality of
[45] A.A. Toropov, A.P. Toropova, E. Benfenati, Additive SMILES-based carcinogenicity
correlation to validate QSAR, SAR QSAR Environ. Res. 30 (2019) 63–80.
models: probabilistic principles in the search for robust predictions, Int. J. Mol. Sci.
[20] P. Kumar, A. Kumar, J. Sindhu, S. Lal, QSAR models for nitrogen containing
10 (2009) 3106–3127.
monophosphonate and bisphosphonate derivatives as human farnesyl

13
P. Kumar, A. Kumar
[46] A.P. Toropova, A.A. Toropov, D. Leszczynska, J. Leszczynski, CORAL and Nano-
QFAR: quantitative feature - activity relationships (QFAR) for bioavailability of
nanoparticles (ZnO, CuO, Co3O4, and TiO2), Ecotoxicol. Environ. Saf. 139 (2017)
404–407.
[47] P.G.R. Achary, A.P. Toropova, A.A. Toropov, Combinations of graph invariants and
attributes of simplified molecular input-line entry system (SMILES) to build up
models for sweetness, Food Res. Int. 122 (2019) 40–46.
[48] M. Duhan, R. Singh, M. Devi, J. Sindhu, R. Bhatia, A. Kumar, P. Kumar, Synthesis,
molecular docking and QSAR study of thiazole clubbed pyrazole hybrid as
α-amylase inhibitor, J. Biomol. Struct. Dyn. (2019), https://doi.org/10.1080/
07391102.2019.17.
[49] A.P. Toropova, A.A. Toropov, QSPR and nano-QSPR: what is the difference? J. Mol.
Struct. 1182 (2019) 141–149.
[50] A.P. Toropova, A.A. Toropov, The index of ideality of correlation: a criterion of
predictability of QSAR models for skin permeability? Sci. Total Environ. 586 (2017)
466–472.
[51] A.A. Toropov, A.P. Toropova, Use of the index of ideality of correlation to improve
predictive potential for biochemical endpoints, Toxicol. Mech. Methods 29 (2019)
43–52.
[52] A.A. Toropov, A.P. Toropova, The Correlation Contradictions Index (CCI): building
up reliable models of mutagenic potential of silver nanoparticles under different
conditions using quasi-SMILES, Sci. Total Environ. 681 (2019) 102–109.
[53] R.B. Aher, K. Roy, Exploring the structural requirements in multiple chemical
scaffolds for the selective inhibition of Plasmodium falciparum calcium-dependent
protein kinase-1 (PfCDPK-1) by 3D-pharmacophore modelling, and docking studies,
SAR QSAR Environ. Res. 28 (2017) 390–414.
[54] S.S. Bhayye, K. Roy, A. Saha, Pharmacophore generation, atom-based 3D-QSAR,
HQSAR and activity cliff analyses of benzothiazine and deazaxanthine derivatives
as dual A2A antagonists/MAOB inhibitors,, SAR QSAR Environ. Res. 27 (3) (2016)
183–202, https://doi.org/10.1080/1062936X.2015.1136840.
[55] A. Golbraikh, A. Tropsha, Beware of q2!, J. Mol. Graph. Model. 20 (2002) 269–276.
[56] L.M. Shi, H. Fang, W. Tong, J. Wu, R. Perkins, R.M. Blair, W.S. Branham, S.L. Dial,
C.L. Moland, D.M. Sheehan, QSAR models using a large diverse set of estrogens,
J. Chem. Inf. Comput. Sci. 41 (2001) 186–195.
[57] G. Schüürmann, R.-U. Ebert, J. Chen, B. Wang, R. Kühne, External validation and
prediction employing the predictive squared correlation coefficient — test set
activity mean vs training set activity mean, J. Chem. Inf. Model. 48 (2008)
2140–2145.
[58] P. Pratim Roy, S. Paul, I. Mitra, K. Roy, On two novel parameters for validation of
predictive QSAR models, Molecules 14 (2009) 1660–1701.
[59] K. Roy, P. Chakraborty, I. Mitra, P.K. Ojha, S. Kar, R.N. Das, Some case studies on
application of “rm2” metrics for judging quality of quantitative structure–activity
relationship predictions: emphasis on scaling of response data, J. Comput. Chem. 34
(2013) 1071–1082.
14
Chemometrics and Intelligent Laboratory Systems 200 (2020) 103982
[60] P. Gramatica, A. Sangion, A historical excursus on the statistical validation
parameters for QSAR models: a clarification concerning metrics and terminology,
J. Chem. Inf. Model. 56 (2016) 1127–1131.
[61] P. Kumar, A. Kumar, Nucleobase sequence based building up of reliable QSAR
models with the index of ideality correlation using Monte Carlo method, J. Biomol.
Struct. Dyn. (2019), https://doi.org/10.1080/07391102.2019.1656109.
[62] P. Kumar, M. Nimbhal, K. Bagri, P. Kumar, The index of ideality of correlation: A
statistical yardstick for better QSAR modeling of glucokinase activators, Struct.
Chem. (2019), https://doi.org/10.1007/s11224-019-01468-w.
[63] P.K. Ojha, I. Mitra, R.N. Das, K. Roy, Further exploring rm2 metrics for validation of
QSPR models, Chemometr. Intell. Lab 107 (2011) 194–205.
[64] A. Rescifina, G. Floresta, A. Marrazzo, C. Parenti, O. Prezzavento, G. Nastasi,
M. Dichiara, E. Amata, Development of a Sigma-2 Receptor affinity filter through a
Monte Carlo based QSAR analysis, Eur. J. Pharmaceut. Sci. 106 (2017) 94–101.
[65] C.A. Lipinski, F. Lombardo, B.W. Dominy, P.J. Feeney, Experimental and
computational approaches to estimate solubility and permeability in drug discovery
and development settings1PII of original article: S0169-409X(96)00423-1. The
article was originally published in Advanced Drug Delivery Reviews 23 (1997)
3–25.1, Adv. Drug Deliv. Rev. 46 (2001) 3–26.
[66] D.A. Griffith, J.R. Hadcock, S.C. Black, P.A. Iredale, P.A. Carpino, P. DaSilva-
Jardine, R. Day, J. DiBrino, R.L. Dow, M.S. Landis, R.E. O’Connor, D.O. Scott,
Discovery of 1-[9-(4-Chlorophenyl)-8-(2-chlorophenyl)-9H-purin-6-yl]-4-ethylami-
nopiperidine-4-carboxylic acid amide hydrochloride (CP-945,598), a novel, potent,
and selective cannabinoid type 1 receptor antagonist, J. Med. Chem. 52 (2009)
234–237.
[67] L.E. Klumpers, C. Roy, G. Ferron, S. Turpault, F. Poitiers, J.-L. Pinquier, J.G.C. van
Hasselt, L. Zuurman, F.A.S. Erwich, J.M.A. van Gerven, Surinabant, a selective
cannabinoid receptor type 1 antagonist, inhibits Δ9-tetrahydrocannabinol-induced
central nervous system and heart rate effects in humans, Br. J. Clin. Pharmacol. 76
(2013) 65–77.
[68] R.J. Chorvat, J. Berbaum, K. Seriacki, J.F. McElroy, JD-5006 and JD-5037:
peripherally restricted (PR) cannabinoid-1 receptor blockers related to SLV-319
(Ibipinabant) as metabolic disorder therapeutics devoid of CNS liabilities, Bioorg.
Med. Chem. Lett 22 (2012) 6173–6180.
[69] O. Trott, A.J. Olson, AutoDock Vina, Improving the speed and accuracy of docking
with a new scoring function, efficient optimization, and multithreading, J. Comput.
Chem. 31 (2010) 455–461.
[70] T. Hua, K. Vemuri, M. Pu, L. Qu, G.W. Han, Y. Wu, S. Zhao, W. Shui, S. Li, A. Korde,
R.B. Laprairie, E.L. Stahl, J.H. Ho, N. Zvonok, H. Zhou, I. Kufareva, B. Wu, Q. Zhao,
M.A. Hanson, L.M. Bohn, A. Makriyannis, R.C. Stevens, Z.J. Liu, Crystal structure of
the human cannabinoid receptor CB1, Cell 167 (2016) 750–762, https://doi.org/
10.1016/j.cell.2016.10.004, e14.