EC

JC
The Journal of Engineering and Exact Sciences – JCEC, Vol. 05 N. 01 (2019)
journal homepage: https://periodicos.ufv.br/ojs/jcec
doi: https: 10.18540/jcecvl5iss1pp0049-0062
OPEN ACCESS – ISSN: 2527-1075

IN-SILICO STUDY FOR PREDICTING THE INHIBITION

CONCENTRATION OF SOME HETEROCYCLIC AND PHENYLIC
COMPOUNDS AS POTENT HERBICIDES USING THE MLR - GFA
APPROACH

ESTUDO IN-SILICO PARA PREDIÇÃO DA CONCENTRAÇÃO DE INIBIÇÃO DE ALGUNS

COMPOSTOS HETEROCÍCLICOS E FENÍLICOS COMO HERBICIDAS POTENTES
UTILIZANDO A ABORDAGEM MLR-GFA

S. TUKUR1,*, G.A. SHALLANGWA1, A. UZAIRU1 and A. IBRAHIM1

Ahmadu Bello University, Department of Chemistry, Zaria, Kaduna State, Nigeria

*
Corresponding author. Ahmadu Bello University, Department of Chemistry, Zaria, Kaduna State, Nigeria, Phone: +2348139475967 e-mail
addressl:saidutukur@gmail.com (S. TUKUR).

ARTICLE INFO ABSTRACT

Article history: The study of the quantitative structure-activity relationship (QSAR) was used in a set of data
Received 2018-12-20
from 43 heterocyclic and phenylic inhibitor compounds in order to establish a correlation
Accepted 2019-01-29
Available online 2019-03-08 between the inhibitory concentrations of the compounds in question and their structures. The
optimization method of the density functional theory (DFT) was used to minimize the energy
k e yw o r ds of the 3D structures using the Becke functional hybrid Exchange (B3) parameter with the
Herbicide Lee, Yang, and Parr Functional Correlation (LYP), commonly called the B3LYP functional
QSAR
Multiple Linear Regression (MLR) Hybrid and 6-31G* Basis Set (B3LYP/6-31G*) method, to discover their molecular
Genetic Function Algorithm (GFA) Quantum descriptors. Five models of QSAR were generated with the technique of genetic
Applicability Domain function algorithm (GFA). Among the five models generated, model 1 was selected as the
Y- Randomization best model because of its statistical significance (Friedman's LOF = 0.3008, R2 = 0.9784,
R2adj = 0.9739, Qcv2 = 0.9675 and R2pred = 0.7348). The meticulous model was evaluated by
means of the Leave One out cross-validation (LOO-CV) approach, external validation of the
compounds of the test set, Y -randomization test and applicability domain (Williams Plot).
The proposed QSAR model was highly predictive and vigorous with good validation
parameters. The molecular descriptors used in the model should be considered of great
importance in improving the inhibitory concentrations of the herbicides and also in the
conception of new herbicides with a higher concentration of inhibitor.

RESUMO
O estudo da relação quantitativa estrutura-atividade (QSAR) foi utilizado em um conjunto
de dados de 43 compostos de inibidores heterocíclicos e fenílicos, a fim de estabelecer uma
correlação entre as concentrações inibitórias dos compostos em questão e seus Estruturas.
O método de otimização da teoria funcional da densidade (DFT) foi utilizado para
minimizar a energia das estruturas 3D utilizando o parâmetro híbrido funcional de Becke
(B3) com a correlação funcional de Lee, Yang e Parr (LYP), comumente denominada
B3LYP funcional híbrido e 6-31G * base Set (B3LYP/6-31G *) método, para descobrir seus
descritores Quantum molecular. Cinco modelos de QSAR foram gerados com a técnica de
algoritmo de função genética (GFA). Entre os cinco modelos gerados, o modelo 1 foi
selecionado como o melhor modelo por causa de sua significância estatística (LOF de
Friedman = 0,3008, R2 = 0,9784, R2adj = 0,9739, Qcv2 = 0,9675 e R2pred = 0,7348). O
modelo meticuloso foi avaliado por meio da abordagem Leave One out Cross-Validation
(LOO-CV), validação externa dos compostos do conjunto de teste, teste de randomização Y
e domínio de aplicabilidade (Williams Plot). O modelo QSAR proposto foi altamente
preditivo e vigoroso com bons parâmetros de validação. Os descritores moleculares
 EC
JC
JCEC - ISSN 2527-1075.

utilizados no modelo devem ser considerados de grande importância na melhoria das

concentrações inibitórias dos herbicidas e também na concepção de novos herbicidas com
maior concentração de inibidores.
 EC
JC
JCEC - ISSN 2527-1075.

1. INTRODUCTION These QSAR methods were established by Hansch and free-

Wilson (Takač et al., 1991) starting from 1D, 2D linear free-
energy correlation, 3D QSAR of Crammer, 4D. Hopfinger, 5D
Herbicides are chemical compounds that inhibit different Vedani and 6D (Hansch et al., 1963). The 1D, 2D and other
metabolic processes in plants. They are extensively used in similar methods were generally referred to as "Classic
agricultural purposes as an agent of destruction against Methods ". The classical method was succeeded with the
undesirable crops (weeds). Weeds usually struggle with other approximation of the genetic function algorithm (GFA)
species of crops for nutrients, sunlight, and water. In addition, (Rasulev et al., 2005). The GFA technique does not only select
weeds are capable of influencing plant growth, causes a decline the integer number of descriptors in the regression study but,
in crop production and quality (Prasad et al., 2011). However, also introduces multiple linear regression (MLR) on the use of
weeds are usually the causative agent of most pathogens, pests, straight-line, higher order polynomials, splines and Gaussian
and viruses, which can contribute extremely to the blowout of (Cho et al., 2001, 18). The algorithm technique was used to
diseases in plant and insect pests. The herbicides being the select the best descriptors in the regression equation. The
safest and the most effective way to control weed can play a genetic function algorithm could help as a valuable procedure
vital role in mechanized farming. The fortification of crops for probing the large probability space with a large number of
continually needs the discovery of new innovative herbicides. descriptors for a small number of molecules. The essence of
However, mechanized and large-scale farming is attainable this research was to develop a predictive QSAR model using
with the help of herbicides. The herbicides play a reliable role 30 compounds as the training set. However, if the developed
in the unified management of weeds (Funar et al., 2017). QSAR models were reasonable enough, it may be used to
predict the inhibition concentration of the unknown
A large species of herbicide were developed early from the compounds. In addition, the prosperous QSAR models will
structure to the genre of actions (Troyer et al., 2001). The first certainly decrease the number of compounds to be synthesized
commercial herbicide that was discovered at the beginning of more especially in pharmaceutical laboratories, thereby,
1948 was 2.4-dichlorophenoxyacetic acid (2.4-D) by making it an easy and affordable technique to reduce cost and
Zimmerman (Zimmerman et al., 1948). In addition, the time duration with regards to developing new drugs and
tendency in its structure and chemical activity was used in the agrochemicals.
designation of new inhibitors (Liu et al., 2013). The herbicides
perform their function by influencing the electron transport
system of the plants during the process of photosynthesis 2. MATERIALS AND METHOD
(Dayan et al., 2012). The slowdown of the electron flow to
participate in the binding of the plastoquinone and the 2.1 Experimental Dataset
eccentricity of the electron in the photosystem I (PSI) was of
great significance in photosynthesis. The relocation of the
43 compounds were used in this study as a data set,
electron from the PSI to its QB binding site of the D1 subunit
comprising a mixture of Sulfonyl Urea, Triazines, Benzoate,
of photosystem II (PSII) is succeeded with the help of the
Acetamides, Pyridines, etc. The inhibition concentration of
plastoquinone. A lot of classes of herbicides (e.g. urea, triazine
these compounds has been reported in the literature (Gand et
pyridines etc.) can be destined to the undiscerning PSII QB
al., 2015). The dataset, along with its chemical abstract service
binding site. PSII plays a vital role in photosynthesis and also
record number (CASRN), IUPAC names, common names, and
in plant growth (Zhang et al. 2014). QB is a complex proteins
various physicochemical properties has been selected from a
sheath, which captures light from the sun and finishes the light
simplified molecular input line String file system Commonly
oxidation of water to atmospheric oxygen. A sequence of light-
abbreviated as SMILES. The result of the inhibitory
inspired electron transfer reactions proceeds in the D1 and D2
concentration was expressed in pLS50 (pLS50 = - log(LS50)
subunit of PSII. The QB binding site of D1 protein is the
which is the concentration at 50% maximum inhibition.
finishing point of the electron transfer process in the PSII and
is also the livelihood target of most herbicides. Herbicides
always displace the plastoquinone from its QB binding position 2.2 Geometry Optimization
(Pfister et al., 1979).
The procedure of calculating the equilibrium geometry or
The quantitative structure-activity relationship (QSAR) is a lowest energy conformation of a molecule is called molecular
technique used in the establishment of mathematical or optimization (Abdulfatai et al., 2015). The 2D structures of the
computational models that strive to establish a good correlation compounds were drawn with the help of ChemDraw Ultra
between molecular structures (descriptors) and properties version 12.0 software and the drawn compounds were saved in
(concentrations) using a Chemometric method (Cruz et al., a format called CDX file format (Ibrahim et al., 2018). The 3D
2014). However, in the context of the molecular design, the conformations of the compounds were optimized using a
word “structure” refers to the molecular descriptor of the programmed software package called Spartan 14.0 VI. 1.4
compounds (Verma et al., 2010). While, the Chemometric wave function (Arthur et al., 2016). The molecular structures
method employs artificial neural networks (ANN), principal were in the first place pre-optimized using the Semiempirical
component regression (PCR), Principal Component Analysis method purposely to reduce the tension in the molecule before
(PCA), multiple linear regression (MLR), partial least squares submitting the whole molecules to DFT (density functional
(PLS), genetic function algorithm (GFA) etc. However, several theory) exploring the Becke three parameters exchange
approaches to the QSAR studies have been established over a functional (B3) hybrid with Lee, Yang and Parr Functional
period of about a hundred years ago and have been abetted as a Correlation (LYP), usually abbreviated as B3LYP (Lee et al.,
good predictive measure, particularly in the design of new 1988) and by the use of 6-31G* basis set for complete
narcotics drugs and in the synthesis of new agrochemicals.
 EC
JC
JCEC - ISSN 2527-1075.

optimization of the 3D structures. The optimized spartan files

of all the molecules were then saved in a format called the SDF 2.7 Relative Importance of Each Descriptor to the Model
file format, which is the only input entry format that can be
recognized by the Padel-Descriptor V 2.20 software (Adeniji et The definitive meaning of the mean effect (ME) of each
al. 2018). descriptor was used as a good measure to evaluate the absolute
prominence and impact of each descriptor to the model. The
2.3 Molecular Descriptor Calculations mean effect can be calculated using Equation 1, given below
as:

∑
The molecular descriptor is a mathematical value or figure

= ( )
that describe properties of molecules acquired from an accurate
∑ ∑
algorithm or experimental procedure (Olasupo et al., 2017).
Quantum chemical descriptors were calculated using the
optimization software of Quantum chemistry known as Spartan
' 14 version 1.1.2 (Abdulfatai et al., 2017). The descriptors of Where, ME is the mean effect of descriptor J in the developed
the various dimensions (1D, 2D, and 3D) were calculated with QSAR model, while βj, is the coefficient of descriptor J, Dj is
the help of Padel software version 2.18. The total molecular the value of each descriptor in the data set for each molecule in
descriptors of 1875 (1444 1D, 2D, and 431 3D) were generated the training compounds, m is the number of descriptors that
from the Padel descriptor software and the Spartan "14 " appear in the model and n is the number of training set
Software (Arthur et al., 2016). compounds (Minovski et al., 2013)

2.4 Data Normalization 2.8 Degree of Contribution of Selected Descriptors

The descriptors obtained from the Padel software were The significant impact of each designated descriptor is to
unvarying by a procedure that makes use of range; Maximum take into account the standardized regression coefficients ( )
and minimum with the dispersion of the molecular descriptors as the factor that must be considered in measuring the degree
using standard deviation or variance. Thus, the sprinkling of of impact of each descriptor to the model. The standardized
the descriptors was achieved before the data were transformed regression coefficients can be calculated using Equation 2,
over into an N distribution (0, 1, 2... N). This procedure given by the expression below:
enables the correlation between descriptors much less
= ( )
redundant (Panchal et al., 2013).

2.5 Data Pretreatment

Where is the regression coefficient of descriptor j. and
The descriptors generated from the PaDEL software were are the standard deviations of each descriptor and
pre-treated using software popularly known by the pharmacist concentration respectively. Emphasis was placed on certain
as DTC LAB, meaning; Drug Theoretical Cheminformatics descriptors that reveal a larger standardized coefficient value.
Laboratory. However, since a large number of descriptors have
been generated, it is therefore important to employ an 2.9 Model development
applicable technique to eliminate the less important
descriptors. Molecular descriptors showing almost equal values The multiple linear regression (MLR) study was
were detached from all ensembles and the molecular performed with the aid of the genetic function algorithm
descriptors having a very low variance were also ignored in (GFA) that was incorporated in the material studio software
order to decrease the Loch descriptor. The descriptors showing purposely to discover the ideal number of the descriptor in the
a small level of intercorrelation were retained (Roy et al., model. Genetic function algorithm is a computer-aided
2015). descriptor calculation method that has been stirred from
engineering and computer scientist (Saidi et al., 2016). GFA
2.6 Generation of Training and Test Set not only inevitably selects the prime number of descriptors in
the regression analysis but, also creates multiple linear
In the development of a validated QSAR model, the data regression models (MLR) by the use of linear, higher order
set of 43 compounds were divided according to Kennard and polynomials, splines and Gaussians (Cho et al., 2001). The
Stones into 70% of training and 30% of the test set using MLR showed an orthodox correlation between the
information division software obtained from the Drug concentrations (pLC50) and the molecular descriptors. The
Theoretical Cheminformatics Laboratory (DTC LAB.) universal expression for MLR is given by equation 3, below:
(Kennard and Stone 1969). The Kennard and Stone algorithm
existed with a prodigious success in the recent QSAR studies = + + + +⋯ ( )
and remained a valuable tool to build both training and test sets
compounds. 70% of the entire data set were used in the Where, Y is the dependent variable (inhibition concentration),
development of the QSAR model. The remaining 30% of the X1, X2, …, Xn are the independent variables (descriptors)
data set were used to validate the developed QSAR model
coefficients of , , …,
existing in the model with their equivalent regression
externally. , respectively, and is the
constant term of the model. The explanation of the impact of
the individual descriptors X1, X2, …, Xn depends on the
 EC
JC
JCEC - ISSN 2527-1075.

corresponding coefficient value of each descriptor in the However, the predictive power of the QSAR model was
regression equation and its numerical sign (Roy et al., 2015). also determined by the use of Friedman's lack of Fit (LOF),
However, each regression coefficient should be significant at which is one among other criteria of internal validation
the 95% probability level (P < 0.05), otherwise called a 95% assessment. The lack of Friedman's Fit (LOF) was calculated
confidence limit. This can be confirmed from a Student 't' test. using Equation 6, given below as:
The number of compounds and the number of descriptors must
+,- = (2)
be in a ratio of at least 5:1. MLR model that fits well the
/ + &#
. − 1
normalized data will give a scatter plot of experimental
inhibition concentration versus predicted inhibition 0
concentration which eventually shows a lower deviation of the
contour points from the line of fit (figure 1). Where SEE is the standard estimation error, also known as
standard deviation (SD), p is the number of independent
2.10 Evaluation of the QSAR Models variable in the model, d is a user-defined smoothing parameter,
c is the number of terms in the model, and N is the number of
The developed QSAR models were evaluated using the the compound in the training set. However, a model is alleged
following statistical parameters, N (number of compounds in to be vigorous if it has a smaller SEE value. The SEE can be
regression), P (number of descriptors), R2 (squared correlation calculated using Equation 7, expressed below as:
coefficient), F- test (Fischer's value), Q2CV (Cross-validation

− #$%&.
correlation coefficient) R2pred. (square correlation coefficient
=3 (6)
%4#.
0−5−
for external tests). The R2 regression coefficient and the Q2CV
cross-validation correlation coefficient are the two most
important determining factors that should be taken into
consideration in accepting any validated QSAR model (Arthur Where, %4#. and #$%&. are the experimental and predicted
et al., 2016). A model is alleged to be significant only if it has inhibition concentration of the compounds in the training set. N
fulfilled the following conditions: R2 > 0.6, Q2CV > 0.6 and is the number of training compounds, P is the number of
R2pred > 0.5. Taking into account the statistical parameters, we descriptors in the model (Jalali et al., 2004). In addition,
opted for a high-value model Q2CV (cross-validation correlation another factor that should be considered highly important when
coefficient) and R2 (correlation coefficient for the training set) accessing the internal validation of a QSAR model is the
as the best model. Leave-One-out cross-validation coefficient. The cross-
. validation regression coefficient (Q2CV) can be calculated using
2.11 Internal Validation of the QSAR Model Equation 8, given by:

∑9 −
7/8 = − (;)
%4#. #$%&.
Internal validation is the first step in validating the

∑9 −:
QSAR model. The expected results of internal validation
indicate that the model permeates a higher level of stability and %4#.

Where <=>?. , <?@=A. , BCD <E are experimental, predicted and the
reliability (Abdulfatai et al., 2016, 1). The Square correlation
coefficient (R2) designates the division of the total variant
accredited to the model. The closer the value of R2 to 1.0; the mean inhibition concentration values of the training set
better the model generated. R2 is one of the most commonly compounds (Jalali et al., 2004).
used internal validation parameters and can be calculated using
Equation 4, expressed below as: 2.12 External validation of the QSAR Model

∑ −
= − ())
! . #$%&.
The developed QSAR model was externally validated to
∑ −
confirm its robustness. Therefore, the external validation of the
! . '$ ( model was evaluated centered on its R2 value for the
compounds in the test set. Therefore, the external predictive
Where, ! . , #$%&. , and '$ ( are observed, predicted a strength and the extrapolation of the models were calculated
mean inhibition concentration of the compounds in the training using the regression coefficient expression given by equation
set (Alho et al., 2010). The R2 value is directly proportional to 9, below:
the number of descriptors in the model. Therefore, we cannot
∑ −
= − (G)
#$%&.'% ' %4#.'% '
only rely on the R2 value in the development of the model.
F% '
∑ − : F$
However, R2 needs to be adjusted to the number of descriptors
used in the model. The adjusted R2 can be calculated using #$%&.'% ' (
equation 5 given below as:
Where, #$%&.'% ' , %4#.'% ' , are the predicted, experimental
− ( − ) −#
= −( − ) = (*) : F$
−#− −# +
inhibition concentration of the compounds in the test set and
&.
( is the mean experimental inhibition concentration of
the training set compounds (Tropsha et al., 2003). In addition,
Where n is the number of training set compounds and p is the predictive capacity of the QSAR model was also calculated

root mean square error ( HIJ ) given by equation 10, below:

the number of descriptors used in the model. using the excluded compounds in the test set with the aid of the
 EC
JC
JCEC - ISSN 2527-1075.

$ = $'% ' K − LM$'% ' − $'% '! LN ( )

In the randomization test, the Multiple Linear Regression
models were completed by scrambling the inhibition
concentration (pLC50), while maintaining the descriptors
Where, $'% ' is the coefficient of determination for test set unchanged. The models that come after the first model usually
compounds and $'% '! is the square correlation coefficient when
have a fundamentally low R2 and cross-validation Q2CV values
for a few numbers of trials, which confirms that the generated
the plot of the experimental inhibition concentration against the models remained predictive and reliable. 10 Y-randomization
predicted inhibition concentration of the test set was plotted at tests were performed and it was perceived from the result that,
zero intercepts (Roy et al., 2012). The values of k and k′ are about four of the models have an estimation of R2 and Q2CV <
the slopes of the regression equation of the predicted inhibition 0.5. This evaluation confirms that the model generated was
concentration versus experimental inhibition concentration and
powerful enough and not acquired by chance occurrence
vice versa (Shen et al., 2002). The k and k′ can be calculated
(Arthur et al., 2016).
using equation 11, given below as:

∑ : ∑ :
2.14 Evaluation of the Applicability Domain of the Model
ķ= &ķ =P
( )
∑: The developed QSAR model was further evaluated by

and : are the predicted, mean experimental

plotting the applicability domain (Williams plot), which is the
Where plot of standardized residual versus leverages. The
concentrations of the training set compounds respectively. applicability domain of a QSAR model is used purposely to
Other computable relationships between the inhibition examine the presence of outliers and influential compounds if
concentration and the descriptors were investigated by means there is and to confess the consistency and robustness of the
of the randomization test, notably known as the Y - generated model (Ibrahim et al., 2018). Leverage is used to
randomization test, which is one among other QSAR external plot the applicability domain of a QSAR model. The leverage
validation methods. The areas of the Y -segment were mixed can be calculated using equation 14, expressed below as:

T = ( )U V , ( = V, … #, ) ( ))
and new QSAR models were developed using the same
arrangement of variables as they existed in the unrandomized F F

is the training set matrix I, X is an × V descriptor

model.
Where
However, we employed the use of a parameter denoted by matrix of the training set, and XT is the transpose matrix X used
Rp2, which incriminate the model R2 for the difference between to generate the model. However, leverage being a prognostic
the mean square correlation coefficient (R2rand.) of the tool; therefore, compounds in the test set or training set with
randomized models and the square correlation coefficient (R2) leverage values hi < h* are considered to be firmly predictive
of the un-randomized model. The Rp2 parameter was calculated by the model. The domain with the reliable predictive ability is
by the mathematical expression given by equation 12, below: that with has compounds (both training and test set) with

residual values not > ± . Therefore, the compounds in the test

leverage values within the threshold (hi < h*) and standardized
# = × M( − &. ) ( )
set or training set with (hi > h*) are believed to be influential
as projected by the model. For the training set, the Williams
Where the parameter R2p ensure that the generated models plot is purposely exploited to identify compounds with the best
were not acquired by coincidence. Meanwhile, we projected structural impact (hi < h*) in the development of the model
that the estimation of R2p should be more prominent than 0.5 (Arthur et al., 2016). The warning leverage (T∗ ) that is the
for a credible model. boundary for the X values can be expressed using equation 15,
given below as:

(V + )
2.13 Y-randomization test
T∗ = ( *)
The Y- randomization test is an important criterion for the
external validation of the developed QSAR model. To
safeguard this, the QSAR model produced was predictive and Where n is the number of compounds in the training set, and k
not acquired by fortuitous, the Y-randomization test was is the number of descriptors used in the model.
employed using the training set compounds, as suggested by
2.15 Quality assurance of the model
the Y randomization test, / # should be greater than 0.5 (/ #
Tropsha (Tropsha et al., 2003). However, for a model to pass

> 0.5). The / # can be calculated using equation 13, expressed The methods of internal and external validation of a QSAR
model were the two most significant techniques used in
below as:
assessing the stability, robustness, reliability and predictive
/ # = ⌈ − ( $) ⌉ ( )
.
capacity of the QSAR model. The validation parameters were
compared with the recommendation standard (Veerasamy et
Where, / # is the coefficient of determination for Y
al., 2011). Table 1.0; outline the Standard General
recommendation values for the internal and external validation
randomization test, is the correlation coefficient for Y- parameters that guarantee whether to accept or to reject a
randomization test and $ is average of the random model model.
(Tropsha et al., 2003).
 EC
JC
JCEC - ISSN 2527-1075.

3. RESULTS AND DISCUSSION

Multiple linear regression and genetic function algorithms Model 1

that were incorporated in a material Studio software were used
to spawn five models. Among the five models generated, pLC50 = 0.037617197 * (AMR) + 0.948665366
model 1 was selected as the best model because of its statistical *(SpMax8_Bhp) - 2.448219792 * (FPSA-2) + 0.133743691
significance (Friedman's LOF = 0.3008, R2 = 0.9784, R2adj = *( MOMI-YZ) + 0.114650757 * (RDF50m) - 0.342918766
0.9739, Q 2cv = 0.9675 and R2pred = 0.7348. However, all the
validation parameters that best endorse the reliability and Friedman’s +,- = 0.3008, '$ ( = 0.9784, & [ '%& =
robustness of the model were reported in table 1. This QSAR 0.9739, \/8 = 0.9675, 0'$ ( = 30, #$%& /'%& . = 0.7348
0'% ' = 13
model is considered highly predictive because it has fulfilled
all the prerequisite necessary for accepting any validated
QSAR model. The IUPAC names of the compounds, the
The validation parameters (internal and external) developed
experimental concentrations, the predicted concentrations and
for each model using the genetic function approximation were
the residual values of the compounds were reported in table 2.
reported in table 3. The molecular descriptors used for the
The low residual values (the difference between the predicted
training and test compounds with their experimental and
and experimental inhibition concentrations) means that the
predicted inhibition concentration were reported in table 4 and
model has an extraordinary predictive capacity. Five molecular
5, respectively.
descriptors were selected to build a linear model which shows
an improved inhibition concentrations result of the compounds
The Pearson correlation matrix of the five descriptors used
from the genetic function algorithm.
in these computational study was reported in table 6. The
.
correlation coefficient between each descriptor in the model
was significantly low, which means that there is no such
correlation between the descriptors used in the development of
the model.

Table 1- General minimum recommended value for an acceptable QSAR model.

Symbol Name Acceptable Value

≥ .2
5G*% < . *
Coefficient of determination

\/8 ≥ .*
Confidence interval at 95% confidence level

− \/8 Difference between `J and abc < .

Cross-validation coefficient
J

0%4'. '% ' %' ≥*

/ # > .*
Minimum number of external test set
The coefficient of determination for Y-randomization

Table 2- Showing Herbicide IUPAC Name, Experimental pIC50, Predicted pIC50 values and Residual Values of the
generated MLR model.
S/N IUPAC Name pLC50Exp pLC50Pred. Residual
.
1 2.96 2.81 -0.15
2

3.04 3.01 -0.03

5

3.26 3.07 -0.19

3.03 2.77 -0.26

10 3.04 2.76 -0.28

 EC
JC
JCEC - ISSN 2527-1075.

11 2.18 2.30 0.12

12 3.32 3.17 -0.15
13 1.8 1.64 -0.16
14 2.02 2.29 0.26
15 0.52 0.46 -0.06
16 5.51 5.34 -0.17

17

3.55 2.37 -1.18

19 4.14 4.43 0.29

20 2.61 2.66 0.05

21 4.98 4.19 -0.79
22 2.61 2.30 -0.31
23 3.18 3.30 0.12
24 5.53 5.56 0.03
25 3.21 3.45 0.24
26 4.16 3.17 -0.99

27 4.2 5.15 0.95

28 4.16 4.26 0.10
29 6.4 6.19 -0.21
30

3.6 3.78 0.18

33 5.85 5.45 -0.40

34

3.11 3.35 0.24

36 1.91 2.42 0.51

37

5.34 5.65 0.31

39 1-chloro-N-(2-chloro-4-fluoro-5-{6-fluoro-1,3-dioxo-hexahydro-1H-pyrrolo[1,2-c]imidazolidin-2- 4.19 3.98 -0.21

yl}phenyl)methanesulfonamide

40 6.04 5.98 -0.06

41 4.33 4.03 -0.30

42 4.44 4.85 0.41

 EC
JC
JCEC - ISSN 2527-1075.

43 5 5.15 0.15

Table 3- Validation parameters for each model using Genetic Function Approximation (GFA)
S/N Model 1 Model 2 Model 3 Model 4 Model 5 Threshold value

1 Friedman LOF 0.300795 0.305711 0.324264 0.326245 0.327771

2 R-squared 0.9784 0.978047 0.976714 0.976572 0.976463
3 Adjusted R-squared 0.9739 0.973473 0.971863 0.971691 0.971559
4 Cross validated R-squared 0.967515 0.967247 0.964517 0.964853 0.960061
5 Significant Regression Yes Yes Yes Yes Yes
6 Significance-of-regression F- 217.4193 213.846 201.3358 200.0839 199.1301
value
7 Critical SOR F-value (95%) 2.64405 2.64405 2.64405 2.64405 2.64405
8 Replicate points 0 0 0 0 0
9 Computed experimental error 0 0 0 0 0
10 Lack-of-fit points 24 24 24 24 24
11 Min expt. error for non- 0.21014 0.21185 0.218184 0.218849 0.219361
significant LOF (95%)

Table 4- Calculated descriptor values for the training sets with their experimental and predicted concentrations.
Molecules AMR SpMax8_Bhp FPSA-2 MOMI-YZ RDF50m Experimental Predicted
pLC50 pLC50
4 77.1585 2.192375 0.971161 1.368516 4.938247 3.04 3.01
5 81.3337 2.246656 0.597721 6.505179 10.84811 5.63 5.50
6 73.6698 1.855351 0.62989 3.441717 9.041087 4.41 4.14
8 77.7131 2.368499 0.990879 1.027995 4.619998 3.26 3.07
9 47.3379 1.147726 0.663697 2.028204 1.060182 0.95 1.29
10 85.9427 2.45754 1.622312 1.648164 11.35794 3.03 2.77
11 61.6751 1.918317 1.000945 2.015574 9.995276 3.04 2.76
12 63.455 2.303188 1.210624 2.76899 5.756426 2.18 2.30
13 75.9645 2.413115 1.603034 3.894527 15.41641 3.32 3.17
14 49.6439 1.400916 0.854616 1.769071 5.600613 1.8 1.64
15 46.9027 0.907671 0.557558 1.631555 10.00544 2.02 2.28
16 18.5507 0.817568 0.413431 2.539778 0.001403 0.52 0.46
18 56.9644 1.303324 0.87168 2.500567 5.616158 1.66 1.88
20 84.3649 2.276436 0.832989 2.848138 9.616859 4.14 4.43
21 98.324 2.699828 1.846134 2.136735 8.547054 2.61 2.66
23 61.3448 1.467804 0.612111 1.200149 2.45737 2.61 2.300
25 71.256 2.257071 0.735835 1.7323 23.1274 5.53 5.56
26 68.3294 2.31343 0.87362 4.708859 4.723874 3.21 3.45
30 133.5519 2.811904 1.9543 3.782769 27.18359 6.4 6.19
32 122.1024 2.768123 1.194735 3.399947 15.81433 5.95 6.22
33 97.112 2.650921 1.659118 1.221423 16.1747 3.6 3.78
34 85.6045 2.349084 0.610816 5.913709 9.17447 5.85 5.45
36 108.0014 2.75524 2.319869 3.399239 19.56198 3.11 3.35
39 91.0298 2.247382 0.490221 4.857742 8.623428 5.34 5.65
42 111.577 2.671396 1.600715 8.795199 20.34854 6.04 5.98
43 92.7797 2.502606 1.116026 2.007745 8.517236 4.33 4.03
 EC
JC
JCEC - ISSN 2527-1075.

44 76.0325 2.45754 0.797488 2.146059 14.51517 4.44 4.85

45 114.8064 2.882648 1.452049 5.385797 11.07858 5 5.15
1 57.8325 1.25614 0.476029 2.03909 5.959689 2.96 2.81
17 107.7082 2.665925 1.421546 1.768612 20.44443 5.51 5.34

Table 5- Calculated descriptors for the test set with their experimental and predicted Activity.
Molecules AMR SpMax8_Bhp FPSA-2 MOMI-YZ RDF50m Activity Predict
2 53.1163 1.246209 0.547225 2.763915 1.450453 2.51 2.03
3 60.7273 1.479485 0.592166 8.461216 3.651049 3.67 3.45
19 54.3212 1.66715 0.850731 2.593675 7.153605 3.55 2.37
22 91.2674 2.665925 1.074292 2.856846 7.112731 4.98 4.19
24 70.4077 1.840996 0.723742 1.335393 7.309965 3.18 3.30
27 98.6457 2.684169 1.884564 5.179453 10.28315 4.16 3.17
28 115.3564 2.665925 1.619995 2.890615 19.21301 4.2 5.15
29 74.3812 1.977655 0.678168 9.250508 3.035912 4.16 4.26
31 125.5065 2.745364 1.025938 5.845965 24.63727 5.71 8.08
35 63.5926 2.315266 1.754143 4.465842 7.913456 2.04 1.46
37 55.904 1.718534 0.723825 3.15644 3.283551 1.91 2.42
38 81.9939 2.45754 0.880807 2.441991 13.95963 4.98 4.84
41 88.4505 2.476088 1.183455 2.169109 10.96011 4.19 3.98

Table 6- Pearson’s correlation matrix of the descriptors in Selected QSAR model.

AMR SpMax8_Bhp FPSA-2 MOMI-YZ RDF50m
AMR 1
SpMax8_Bhp 0.890651 1
FPSA-2 0.694943 0.699339 1
MOMI-YZ 0.378794 0.333113 0.089971 1
RDF50m 0.719901 0.639811 0.605965 0.256842 1

Table 7- Y-Randomization Test Parameters

Model R R^2 Q^2
The Y-randomization test being one among other
Original 0.989141 0.9784 0.967515 essential methods used in the validation of the QSAR model
Random 1 0.522795 0.273315 -0.16266 externally. The Y- randomization test was reported in Table

and \/8 for the number of trials allow us to say that, the
Random 2 0.500953 0.250954 -0.13068 7. It was observed from the result that, the low values of
Random 3 0.46971 0.220627 -0.16702
/ # value was much higher than 0.5 which signifies that,
constructed model was reliable, stable and robust. While, the
Random 4 0.579175 0.335444 -0.09836
Random 5 0.354832 0.125906 -0.29534 the built model was powerful and not inferred by Trial and
Random 6 0.314135 0.098681 -0.47183 error.
Random 7 0.559705 0.31327 -0.13038
The description and other related statistical parameters
Random 8 0.306126 0.093713 -0.41786 that may have a greater impact on the selected descriptors
Random 9 0.262277 0.068789 -0.70639 were reported in table 8. The appearance of 2D and 3D
Random 10 0.360066 0.129647 -0.54853 descriptors in the model shows that these types of
descriptors will have a greater affinity (increase the
inhibition concentration of the compounds). The values of
Random Models Parameters the variance inflation factor (VIF) for all the five descriptors
Average r : 0.422977 of the model were lower than 7, which indicates that the
descriptors were orthogonal and the model validity was
Average r^2 : 0.191035 highly significant. The null hypothesis suggesting that there
Average Q^2 : -0.31291 is no significant relationship between the inhibition
cRp^2 : 0.884432 concentration and the descriptor used in the construction of
the model at P < 0.05. The P- values of the descriptors in the
 EC
JC
JCEC - ISSN 2527-1075.

model at 95% confidence limit were reported in table 8, which

are all less than 0.05. This implies that the null hypothesis was standardized residual values was virtually on both sides of the
actually rejected. Thus, we accepted the alternative hypothesis zero (Jalali et Al., 2004).
which predicts that there exists a significant relationship
between the concentration of the inhibitor molecule and the The leverage values for the complete set of compounds in
descriptor. Thus, we can conclude that there is a significant the data set were plotted against their standardized residual
relationship between the concentration of the inhibitor values. The Williams plot, that is, a plot of standardized
compounds and the descriptors used in the construction of the residual versus leverage value was shown in Figure 4. This plot
model at 95% probability level. really helps us to inspect outliers and influential compounds in
the model. From our result, it is clear that all compounds
The plot of the predicted inhibition concentration against (training and testing) were within the square discrimination of
the experimental concentration of inhibition of the training and ± 3 of standardized residual. Therefore, no outlier compound is
test compounds was reported in Figure 1 and 2, respectively. found in this study. However, only two compounds from the
The R2 value of 0.9784 for the training set and the R2 value of test set with ID-31 and ID-35 are said to be an influential

than the warning leverage ( T > T∗ ). This behavior was

0.7348 for the test compounds as reported in this study, was in compound. This is because their leverage values are greater
agreement with the genetic function approximation
recommendations report reported in table 1. This confirms the attributed due to differences in their molecular structures as
robustness and reliability of the model. The plot of the compared to other compounds in the entire data set.
standardized residual versus experimental inhibition
concentration shown in Figure 3 indicates that there was no
reasonable error in the generated model since the spread of the

Table 8 - List of the descriptors, their description, classes, and their statistical significance.
Statistics
S/N Descriptor Description Class VIF ME P-Value
Symbol

1 AMR Molar 2D 6.31 0.74 8.54E-08

refractivity

2 SpMax8_Bhp The largest 2D 5.18 0.50 4.15E-05

absolute
eigenvalue of
Burden modified
matrix - n 8 /
weighted by
relative
polarizabilities
3 FPSA-2 PPSA-2 / total 3D 2.33 -0.64 4.03E-14
molecular surface
area

4 MOMI-YZ Y/Z 3D 1.26 0.10 0.00024

5 RDF50m Radial 3D 2.18 0.30 4.19E-10

distribution
function - 050 /
weighted by
relative mass
 EC
JC
JCEC - ISSN 2527-1075.

6 R² = 0.9784

5
Predicted (pLC50)
4

0
0 1 2 3 4 5 6 7
Experimental (pLC50)

Figure 1- A plot of Predicted (pLC50) versus Experimental (pLC50) of the training set

9
8 R² = 0.7348
7
Predicted (pLC50)

6
5
4
3
2
1
0
0 1 2 3 4 5 6
Experimental (pLC50)

Figure 2- A plot of Predicted (pLC50) versus Experimental (pLC50) of the test set.
 EC
JC
JCEC - ISSN 2527-1075.

ℎ∗ = 0.6
training trainig
4 4
set set
3

Standardied Residual
3 test set test set
ID-31
2 2
Residual (pS)

1 1
0 0
ID-35
-1 -1
-2 -2
-3 -3
-4 -4
0 5 0 0,5 1
Experimental (pLC50) leverage

Figure 3- A plot of Residual versus Experimental (pLC50) Figure 4- Williams Plot, A plot of standardized residual
of the training and test set versus Leverage

4. CONCLUSION ADENIJI, S. E.; UBA, S.; UZAIRU, A. A Novel QSAR Model

for the Evaluation and Prediction of (E)-N’-
Benzylideneisonicotinohydrazide Derivatives as the
QSAR studies were carried out in a set of data from 43
Potent Anti-mycobacterium Tuberculosis Antibodies
heterocyclic and phenylic compounds used as potent
Using Genetic Function Approach. Physical
herbicides. The Density Functional Theory (DFT) optimization
Chemistry Research, v.6, n. 3, p. 479-492, 2018.
method was used in minimizing the energy of the 3D
ALHO, J. S.; VÄLIMÄKI, K.; MERILÄ, J. Rhh: an R
structures. The Genetic function algorithm (GFA) that was
extension for estimating multilocus heterozygosity
incorporated into the material studio software was used to
and heterozygosity–heterozygosity correlation.
develop five QSAR models. Among the five models generated,
Molecular Ecology Resources, v.10, n. 4, p. 720-
model 1 was selected as the best model because of its statistical
722, 2010.
significance (Friedman's LOF = 0.3008, R2 = 0.9784, R2adj =
ARTHUR, D. E.; UZAIRU, A.; MAMZA, P.; ABECHI, E.;
0.9739, Qcv2 = 0.9675 and R2pred = 0.7348). The relative impact
SHALLANGWA, G. In Silico Modelling of
of each descriptor on the model was accessed using the mean
Cytotoxic Behaviour of Anti-Leukemic Compounds
effect (ME). However, it is evident that a decrease in the
on HL-60 Cell Line. Journal of the Turkish
molecular surface area PPSA-2/total of the descriptor FPSA-2
Chemical Society, Section A: Chemistry, v.3, n. 2,
with the negative mean effect of - 0.6429 will increase the
p. 147-158, 2016.
inhibition concentration of the compounds. Whereas, an
ARTHUR, D. E.; UZAIRU, A.; MAMZA, P.; ABECHI, S.
increase in AMR (molar refractivity), SPMAX8_BHP
Quantitative structure-activity relationship study on
(Largest absolute eigenvalue of Burden modified matrix - n 8 /
potent anticancer compounds against MOLT-4 and
weighted by relative polarizabilities), Momi-YZ (Y/Z) and
P388 leukemia cell lines. Journal of Advanced
RDF50m (Radial distribution function - 050 / weighted by
Research, v.7, n. 5, p. 823-837, 2016.
relative mass) of the descriptors with positive mean effects of
ARTHUR, D. E.; UZAIRU, A.; MAMZA, P.; ABECHI, E.;
0.7402, 0.5006, 0.9992 and 0.3022 will increase the inhibition
SHALLANGWA, G. Insilco study on the toxicity of
concentration of these compounds. The QSAR result
anti-cancer compounds tested against MOLT-4 and
developed in this study actually correlates with the standard
p388 cell lines using GA-MLR technique. Beni-Suef
recommendation values reported earlier in table 1. The positive
University Journal of Basic and Applied Sciences,
mean effects values of some descriptors actually provide a
v.5, n. 4, p. 320-333, 2016.
good idea in designing new herbicides with improving activity.
BECKE, A. D. Becke's three-parameter hybrid method using
the LYP correlation functional. Journal of Chemical
RE F E R E N C E S Physics, v. 98, p. 5648-5652, 1993.
CHO, D. H.; LEE, S. K.; KIM, B. T.; NO, K. T. Quantitative
structure-activity relationship (QSAR) study of new
ABDULFATAI, U.; UZAIRU, A.; UBA, S. Investigation of fluorovinyloxyacetamides. Bulletin of the Korean
some acetamido-N-benzylacetamide derivatives as Chemical Society, v.22, n. 4, p. 388-394, 2001.
potent anti-convulsant inhibitors. Journal of CRUZ, V. L.; MARTINEZ, S.; RAMOS, J.; MARTINEZ-
Computational Methods in Molecular Design, v.5, SALAZAR, J. 3D-QSAR as a tool for understanding
n. 4, p. 77-83, 2015. and improving single-site polymerization catalysts. a
ABDULFATAI, U.; UZAIRU, A.; UBA, S. Quantitative review. Organometallics, v.33, n. 12, p. 2944-2959,
structure-activity relationship and molecular docking 2014.
studies of a series of quinazolinonyl analogs as DAYAN, F. E.; ZACCARO, M.; LETICIA DE, M.
inhibitors of gamma-aminobutyric acid Chlorophyll fluorescence as a marker for herbicide
aminotransferase. Journal of Advanced Research,
v.8, n. 1, p. 33-43, 2017.
 EC
JC
JCEC - ISSN 2527-1075.

mechanisms of action. Pesticide Biochemistry and RASULEV, B. F.; ABDULLAEV, N. D.; SYROV, V. N.;
Physiology, v.102, n. 3, p. 189-197, 2012. LESZCZYNSKI, J. A Quantitative Structure‐Activity
FUNAR-TIMOFEI, S.; BOROTA, A.; CRISAN, L. Combined Relationship (QSAR) Study of the Antioxidant
molecular docking and QSAR study of fused Activity of Flavonoids. QSAR & Combinatorial
heterocyclic herbicide inhibitors of D1 protein in Science, v.24, n. 9, p. 1056-1065, 2005.
photosystem II of plants. Molecular diversity, v.21, ROY, K.; KAR, S.; Das, R. N. Understanding the basics of
n. 2, p. 437-454, 2017. QSAR for applications in pharmaceutical sciences and
GANDY, M. N.; CORRAL, M. G.; MYLNE, J. S.; STUBBS, risk assessment. Academic Press, 2015.
K. A. An interactive database to explore herbicide SAIDI, A.; MIRZAEI, M. Prediction of AHAS inhibition by
physicochemical properties. Organic & sulfonylurea herbicides using a genetic algorithm and
biomolecular chemistry, v,13, n. 20, p. 5586-5590, artificial neural network. 2016.
2015. SHEN, M.; LET IRAN, A.; XIAO, Y.; GALBRAITH, A.;
HANSCH, C.; MUIR, R. M.; FUJITA, T.; MALONEY, P. P.; KOHN, H.; TROPSHA, A. Quantitative structure-
GEIGER, F.; STREICH, M. The correlation of activity relationship analysis of functionalized amino
biological activity of plant growth regulators and acid anticonvulsant agents using k nearest neighbor
chloromycetin derivatives with Hammett constants and simulated annealing PLS methods. Journal of
and partition coefficients. Journal of the American medicinal chemistry, v.45, n. 13, p. 2811-2823,
Chemical Society, v.85, n. 18, p. 2817-2824, 1963. 2002.
IBRAHIM, M. T.; UZAIRU, A.; SHALLANGWA, G. A.; TAKAČ, M. J.; MEDIĆ-ŠARIČ, M. QSPR, and QSAR in
IBRAHIM, A. In-silico studies of some oxadiazoles Pharmacy. I. Classic QSAR models. Hansch and Fred
derivatives as anti-diabetic compounds. Journal of Wilson's model. Farmaceutski glasnik: glacial
King Saud University-Science, p. 2018. Hrvatskog farmaceutskog društva, v.47, n. 6, p.
JALALI-HERAVI, M.; KYANI, A. "Use of computer-assisted 161-178, 1991.
methods for the modeling of the retention time of a TROPSHA, A.; GRAMATICA, P.; GOMBAR, V. K. The
variety of volatile organic compounds: a PCA-MLR- importance of being earnest: validation is the absolute
ANN approach. Journal of chemical information essential for successful application and interpretation
and computer sciences, v.44 n. 4, p. 1328-1335, of QSPR models. QSAR & Combinatorial Science,
2004. v.22, n. 1, p. 69-77, 2003.
KENNARD, R. W.; STONE, LARRY, A. Computer-aided TROYER, J. R. In the beginning: the multiple discoveries of
design of experiments. Technometrics, v.11, n. 1, p. the first hormone herbicides. Weed Science, v.49, n.
137-148, 1969. 2, p. 290-297, 2001.
LEE, C.; Yang, W.; PARR, R.G. Becke's three-parameter VEERASAMY, R.; RAJAK, H.; JAIN, A.; SIVADASAN, S.;
hybrid method using the LYP. Phys. Rev. B, v.37, p. VARGHESE, C. P.; AGRAWAL, R. K. Validation of
785, 1988. QSAR models-strategies and importance.
LIU, Y.; Zhao, H.; WANG, Z.; Li, Y.; Song, H.; RICHES, H.; International Journal of Drug Design & Discovery,
BEATTIE, D.; Gu, Y.; WANG, Q. The discovery of v.3, p. 511-519, 2011.
3-(1-aminoethylidene) quinoline-2, 4 (1H, 3H)-dione VERMA, J.; KHEDKAR, V. M.; COUTINHO, E. C. "3D-
derivatives as novel PSII electron transport inhibitors. QSAR in drug design-a review. Current topics in
Molecular diversity, v.17, n. 4, p. 701-710, 2013. medicinal chemistry, v.10, n. 1, p. 95-115, 2010.
MINOVSKI, N.; ŽUPERL, Š.; DRGAN, V.; NOVIČ, M. ZHANG, C.; CHANG, S.; TIAN, X.; Tian, Y. 3D-QSAR and
Assessment of applicability domain for multivariate docking modeling study of 1, 3, 5-triazine derivatives
counter-propagation artificial neural network as PSII electron transport inhibitor. Asian Journal of
predictive models by minimum Euclidean distance Chemistry, v.26, n. 1, p. 264, 2014.
space analysis: A case study. Analytica chimica acta, ZIMMERMAN, P. W.; HITCHCOCK, A. E. Plant hormones.
v.759, p. 28-42, 2013. Annual review of biochemistry, v.17, n. 1, p. 601-
OLASUPO, S. B.; UZAIRU, A.; SAGA GIS, B. S. Density 626, 1948.
Functional Theory (B3LYP/6-31G*) Study of
Toxicity of Polychlorinated Dibenzofurans., 2017.
PANCHAL, J. H.; KALIDINDI, S. R.; MCDOWELL, D. L.
Key computational modeling issues in integrated
computational materials engineering. Computer-
Aided Design, v.45, n. 1, 4-25, 2013.
PFISTER, K.; ARNTZEN, C. J. The mode of action of
photosystem II-specific inhibitors in herbicide-
resistant weed biotypes. Zeitschrift für
Naturforschung C, v.34, n. 11, p. 996-1009, 1979.
PRASAD, R. K.; SHARMA, R. 2D QSAR Analysis of
pyrazine carboxamide derivatives as an herbicidal
agent. Journal of Computational Method &
Molecular Design, v.1, p. 7-13, 2011.