Prediction Acidity Constant Using QSPR Models Bull. Korean Chem. Soc. 2005, Vol. 26, No.

12 2007

Prediction Acidity Constant of Various Benzoic Acids and Phenols

in Water Using Linear and Nonlinear QSPR Models
Aziz Habibi-Yangjeh,* Mohammad Danandeh-Jenagharad, and Mahdi Nooshyar
Department of Chemistry, Faculty of Science, University of Mohaghegh Ardebili, P.O. Box 179, Ardebil, Iran
*
E-mail: habibiyangjeh@yahoo.com; ahabibi@uma.ac.ir
Received April 18, 2005

An artificial neural network (ANN) is successfully presented for prediction acidity constant (pKa) of various
benzoic acids and phenols with diverse chemical structures using a nonlinear quantitative structure-property
relationship. A three-layered feed forward ANN with back-propagation of error was generated using six
molecular descriptors appearing in the multi-parameter linear regression (MLR) model. The polarizability term
(π I), most positive charge of acidic hydrogen atom (q+), molecular weight (MW), most negative charge of the
acidic oxygen atom (q−), the hydrogen-bond accepting ability (ε B) and partial charge weighted topological
electronic (PCWTE) descriptors are inputs and its output is pKa. It was found that properly selected and trained
neural network with 205 compounds could fairly represent dependence of the acidity constant on molecular
descriptors. For evaluation of the predictive power of the generated ANN, an optimized network was applied
for prediction pKa values of 37 compounds in the prediction set, which were not used in the optimization
procedure. Squared correlation coefficient (R2) and root mean square error (RMSE) of 0.9147 and 0.9388 for
prediction set by the MLR model should be compared with the values of 0.9939 and 0.2575 by the ANN model.
These improvements are due to the fact that acidity constant of benzoic acids and phenols in water shows
nonlinear correlations with the molecular descriptors.
Key Words : Quantitative structure-property relationship, Artificial neural networks, Acidity constant, Phe-
nols, Benzoic acids

Introduction activity and descriptors is considered. If there is not a well-

defined linear relationship, the discussed method cannot
The macroscopic (bulk) activities/properties of chemical give a perfect QSPR/QSAR model. Artificial neural net-
compounds clearly depend on their microscopic (structural) works (ANNs) are capable of recognizing highly nonlinear
characteristics. Development of quantitative structure-pro- relationships. ANNs are biologically inspired computer
17-20

perty/activity relationships (QSPR/QSAR) on theoretical programs designed to simulate the way in which the human
descriptors is a powerful tool not only for prediction of the brain processes information. ANNs gather their knowledge
chemical, physical and biological properties/activities of by detecting the patterns and relationships in data and
compounds, but also for deeper understanding of the de- learned (or trained) through experience, not from program-
tailed mechanisms of interactions in complex systems that ming. There are many types of neural networks designed by
predetermine these properties/activities. QSPR/QSAR
1-10
now and new ones are invented every week. The behavior
20

models are essentially calibration models in which the of a neural network is determined by transfer functions of its
independent variables are molecular descriptors that describe neurons, by learning rule, and by the architecture itself. An
the structure of the molecules and the dependent variable is ANN is formed from artificial neuron or processing
the property or activity of interest. Since these theoretical elements (PE), connected with coefficients (weights), which
descriptors are determined solely from computational constitute the neural structure and are organized in layers.
methods, a priori predictions of the properties/activities of The first layer is termed the input layer, and the last layer is
compounds are possible, no laboratory measurements are the output layer. The layers of neurons between the input and
needed thus saving time, space, materials, equipment and output layers are called hidden layers. Neural networks do
alleviating safety (toxicity) and disposal concerns. An enor- not need on explicit formulation of the mathematical or
mous number of descriptors have been used by researchers physical relationships of the handled problem. These give
to increase the ability to correlate biological, chemical and ANNs an advantage over traditional fitting methods for
physical properties. To obtain a significant correlation, it is some chemical application. For these reason in recent years,
crucial that appropriate descriptors be employed. 11,12
ANNs have been used to a wide variety of chemical
Various methods for constructing QSPR/QSAR models problems such as simulation of mass spectra, ion interaction
have been used including multi-parameter linear regression chromatography, aqueous solubility and partition coeffi-
(MLR), principal component analysis (PCA) and partial cient, simulation of nuclear magnetic resonance spectra,
least-squares regression (PLS). In some cases, it is more
13-16
prediction of bioconcentration factor, solvent effects on
convenient that a linear relationship between property/ reaction rate and prediction of normalized polarity parameter
2008 Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 Aziz Habibi-Yangjeh et al.
in mixed solvent systems. 21-36
molecular electronic descriptors were calculated by Dragon
It has been shown that the acid-base properties affect the package version 2.1. For this propose the output of the
48

toxicity, chromatographic retention behavior and pharma- HyperChem software for each compound feed into the
ceutical properties of organic acids and bases. On the
37,38
Dragon program and the descriptors were calculated. As a
other hand, interpretation and prediction of pKa values for result, a total of 18 theoretical descriptors were calculated
chemical compounds are of general importance and useful- for each compound in the data sets (242 compounds).
ness for chemists. Although in the last years several
39
Linear correlations. Acidity constant of benzoic acids
theoretical studies have been performed for correlation of and phenols are literature values at 25 ºC. MLR model was 49

pKa values with molecular parameters, but in these studies developed for prediction of pKa values by molecular
linear equations have been used. 38-46
descriptors. The method of stepwise multi-parameter linear
The main aim of present work is to develop a linear and regression was used to select the most important descriptors
nonlinear QSPR models based on molecular descriptors for and to calculate the coefficients relating the pKa to the
prediction pKa values of various benzoic acids and phenols descriptors. The MLR models were generated using spss/pc
with diverse chemical structures (including 242 com- software package release 9.0.
pounds). Neural network generation. The specification of a
typical neural network model requires the choice of the type
Theory of inputs, the number of hidden layers, the number of
neurons in each hidden layer and the connection structure
A detailed description of theory behind a neural network between the inputs and the output layers. The number of
has been adequately described by different researchers. 17-19
input nodes in the ANNs was equal to the number of
There are many types of neural network architectures, but molecular descriptors in the MLR model. A three-layer
the type that has been most useful for QSAR/QSPR studies network with a sigmoidal transfer function was designed.
is the multilayer feed - forward network with back-propa- The initial weights were randomly selected between 0 and 1.
gation (BP) learning rule. The number of neurons in the
20
Before training, the input and output values were normalized
input and output layers are defined by system's properties. between 0.1 and 0.9. The optimization of the weights and
The number of neurons in the hidden layer could be biases was carried out according to the resilient back-
considered as an adjustable parameter, which should be propagation algorithm. The data set was randomly divided
50

optimized. The input layer receives the experimental or into three groups: a training set, a validation set and a
theoretical information. The output layer produces the prediction set consisting of 168, 37 and 37 molecules,
calculated values of dependent variable. The use of ANNs respectively. The training and validation sets were used for
consists of two steps: “training” and “prediction”. In the the model generation and the prediction set was used for
training phase the optimum structure, weight coefficients evaluation of the generated model, because a prediction set
and biases are searched for. These parameters are found is a better estimator of the ANN generalization ability than a
from training and validation data sets. After the training validation (monitoring) set. 51

phase, the trained network can be used to predict (or The performances of training, validation and prediction of
calculate) the outputs from a set of inputs. ANNs allow one ANNs are evaluated by the mean percentage deviation
to estimate relationships between input variables and one or (MPD) and root-mean square error (RMSE), which are
several output dependent variables. The ANN reads the defined as follows:
input and target values in the training data set and changes
the values of the weighted links to reduce the difference N P exp – P cal )
MPD = ---1- ∑ (-----------------------------
i i - (1)
between the calculated output and target values. The error N i=1 exp
Pi
between output and target values is minimized across many
training cycles until network reaches specified level of cal 2
accuracy. If a network is left to train for too long, however, it P
N ( i – i )
exp
P
will overtrain and will lose the ability to generalize. 22-36
RMSE = ∑
i=1
--------------------------------
N (2)

Experimental Section where Pi and Pical are experimental and calculated values
exp

of pKa with the models and N denote the number of data

Descriptor generation. The derivation of theoretical points.
molecular descriptors proceeds from the chemical structure Individual percent deviation (IPD) is defined as follows:
of the compounds. In order to calculate the theoretical calc
descriptors, the z-matrices (molecular models) were con- ⎛ i P--------------------------
– Pexp i -⎞
structed with the aid of HyperChem 7.0 and molecular IPD = 100 × ⎜ ⎟ (3)
⎝ Pi exp
⎠
structures were optimized using AM1 algorithm. In order
47

to calculate some of theoretical descriptors, the molecular The processing of the data was carried using Matlab 6.5. 52

geometries of molecules were further optimized with the The neural networks were implemented using Neural
same algorithm in MOPAC program version 6.0. The other Network Toolbox Ver. 4.0 for Matlab. 50
Prediction Acidity Constant Using QSPR Models Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 2009

Results and Discussion tors. Acidity constant of the compounds decrease with
53

increasing ε and PCWTE descriptors, because basicity of

B

A major challenge in the development of MLR equations phenolic oxygen atom increases with increasing these
is connected with the possible multicollinearity of molecular descriptors. Effects of π , q and MW on pKa are higher than
I
+

descriptors. In order to decrease the redundancy existed in that of the other descriptors, because standardized coeffi-
the descriptors data matrix, the correlation of descriptors cients of π , q and MW are higher than those of the other
I
+

with each other and with pKa of the compounds was descriptors.
examined and collinear descriptors were detected (r > 0.85). The calculated values of pKa for the compounds in
Among the collinear descriptors, one with the lowest training, validation and prediction sets using the MLR model
correlation with the property was removed from the data have been plotted versus the experimental values of it in
matrix. Table 1 demonstrates that all of the descriptors are Figure 1.
strongly orthogonal which reflects the statistical reliability The next step in this work was generation of the ANN
of the model. model. There are no rigorous theoretical principles for
Multi-parameter linear correlation of pKa values of 168 choosing the proper network topology; so different struc-
benzoic acids and phenols versus the molecular descriptors tures were tested in order to obtain the optimal hidden
in the training set gives the results in Table 2. It can be seen neurons and training cycles. Before training the network,
36

from this table that six descriptors are appeared in the MLR the number of nodes in the hidden layer was optimized. In
model. These descriptors are: polarizability index (π ), most I order to optimize the number of nodes in the hidden layer,
positive charge of acidic hydrogen atom (q ), molecular +
several training sessions were conducted with different
weight (MW), most negative charge of acidic oxygen atom numbers of hidden nodes. The root mean squared error of
(q−), the hydrogen-bond accepting ability (ε ) and partial B training (RMSET) and validation (RMSEV) sets were
charge weighted topological electronic (PCWTE) descrip- obtained at various iterations for different number of neu-
tors.
The negative coefficient for π , q , q− and MW descriptors
I
+

indicate that with increasing these descriptors, acidity con-

stant (Ka) increases. With increasing q and q− of the com-
+

pounds, interactions of water molecules with acidic hydro-

gen and oxygen of the compounds increase, then acidic
hydrogen can be easily removed from the compounds.
Polarizability and then the dipole-induced dipole interac-
tions increase with increasing π and MW, as a result acidity
I

of the compounds increases with increasing these descrip-

Table . Correlation coefficients between various theoretical

1

descriptors that have been used in the multi-parameter linear

regression (MLR) and artificial neural network (ANN) models
Descriptor πI q+ q− εB MW PCWTE
πI 1 0.530 0.042 0.150 0.329 0.285
q+ 0.530 1 0.642 0.546 0.368 0.070
q− 0.042 0.642 1 0.236 0.155 0.018
εB 0.150 0.546 0.236 1 0.248 0.038
MW 0.329 0.368 0.155 0.248 1 0.237 Figure 1 . Plot of the calculated values of pKa from the MLR model
PCWTE 0.285 0.070 0.018 0.038 0.237 1 versus the experimental values of it for training, validation and
prediction sets.

Table 2. Descriptors, symbols and results of the multi-parameter linear regression (MLR) modela
No. Descriptor Symbol Coefficient β
1 polarizability term πI −8.3610 0.080
2 most positive charge of acidic hydrogen atom q+ −110.4710 0.521
3 molecular weight MW −0.0051 0.074
4 most negative charge of the phenolic oxygen atom q_ −26.3940 0.321
5 the hydrogen-bond accepting ability εB 34.4450 0.080
6 partial charge weighted topological electronic PCWTE 0.0902 0.101
7 constant 42.2780
a
The β is standardized coefficient of descriptors. The polarizability term (π ) is obtained by dividing the polarizability volume by the molecular volume.
I

The ε is equal 0.3–0.01(E –E ), in which E and E are referring to the LUMO energy for water and HOMO energy for the compound, respectively.
B lw h lw h
2010 Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 Aziz Habibi-Yangjeh et al.
iterations are stopped when overtraining begins. To control
the overtraining of the network during the training proce-
dure, the values of RMSET and RMSEV were calculated
and recorded to monitor the extent of the learning in various
iterations. Results obtained showed that after 77000 iter-
ations the value of RMSEV started to increase very little and
overfitting slightly began (Figure 3).
The generated ANN was then trained using the training
and validation sets for the optimization of the weights and
biases. For the evaluation of the predictive power of the
generated ANN, an optimized network was applied for
prediction of the p a values of the compounds in the
K

prediction set, which were not used in the modeling proce-

dure (Table 3). The calculated values of p a for the K

compounds in training, validation and prediction sets using

the ANN model have been plotted versusthe experimental
values of it in Figure 4.
As expected, the calculated values of p a are in good
K

agreement with those of the experimental values. The

Figure 2. Plot of RMSE for training and validation sets versus the correlation equation for all of the calculated values of p a K

number of nodes in hidden layer. from the ANN model and the experimental values is as
follows:
rons at the hidden layer and the minimum value of RMSEV p a(cal) = 0.99299 p a(exp) + 0.04454 (4)
was recorded as the optimum value. Plot of RMSET and K K

RMSEV versusthe number of nodes in the hidden layer has (R = 0.9931; MPD = 4.5044; RMSE = 0.2648; F = 34295.94)
2

been shown in Figure 2. It is clear that the twenty-four nodes Similarly, the correlation of p a (cal) p a (exp)
in hidden layer is optimum value. K versus

values in the prediction set gives equation (5):

K

The six descriptors appearing in the MLR model (includ-

ing π I, q+, MW, q−, ε B, and PCWTE descriptors) were p a(cal) = 1.01212p a(exp) – 0.08200
K K (5)
considered as inputs for developing the ANN. Then an ANN (R2 = 0.9939; MPD = 5.0361; RMSE = 0.2575; F = 5718.11)
with architecture 6-24-1 was generated. It is note worthy that
training of the network was stopped when the RMSEV Plot of IPD for p a values in the prediction set
K the
versus

started to increases . when overtraining begins. The over-

i.e experimental values of it has been illustrated in Figure 5. As
training causes the ANN to loose its prediction power.34,36 can be seen, the model did not show proportional and
Therefore, during training of the networks, it is desirable that systematic error, because the slope (a = 1.01212) and
intercept (b = −0.08200) of the correlation equation are not
significantly different from unity and zero, respectively and
the propagation of errors in both sides of zero is random
(Figure 5).
Table 4 compares the results obtained using the MLR and
ANN models. The squared correlation coefficient (R2) and
RMSE of the models for total, training, validation and
prediction sets show potential of the ANN model for predic-
tion of p a values of various benzoic acids and phenols in
K

water with one model.

As a result, it was found that properly selected and trained
neural network could fairly represent dependence of the
acidity constant of benzoic acids and phenols in water on the
molecular descriptors. Then the optimized neural network
could simulate the complicated nonlinear relationship
between p a values and the molecular descriptors. It can be
K

seen from Table 4 that although the parameters appearing in

the MLR model are used as inputs for the generated ANN,
the statistics is shown a large improvement. These improve-
Figure 3. Plot of RMSE for training and validation sets versus the ments are due to the fact that p a values of the compounds
K

number of iterations. show nonlinear correlations with the molecular descriptors.

Prediction Acidity Constant Using QSPR Models Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 2011

Table 3. Experimental and calculated values of pKa for various benzoic acids and phenols in water at 25 ºC for training, validation and
prediction sets by multi-parameter linear regression (MLR) and artificial neural network (ANN) models along with individual percent
deviation (IPD)a
No. Compound Exp. MLR IPDMLR ANN IPDANN
Training set

1 2-acetylphenol 9.19 10.867 18.25 9.272 0.89

2 4-acetylphenol 8.05 9.05 12.42 8.794 9.24
3 2-allylphenol 10.28 9.637 −6.25 9.972 −3.00
4 4-bromophenol 9.34 8.42 −9.85 9.126 −2.29
5 2,6-di- -butyl-4-bromophenol
tert 10.83 10.011 −7.56 10.975 1.34
6 2,6-di- -butyl-4-methylphenol
tert 12.23 11.201 −8.41 11.983 −2.02
7 2,6-di- -butyl-4-methoxyphenol
tert 12.15 11.484 −5.48 11.936 −1.76
8 2- -butylphenol
tert 11.24 10.96 −2.49 10.752 −4.34
9 3- -butylphenol
tert 10.1 10.773 6.66 10.504 4.00
10 4- -butylphenol
tert 10.31 10.768 4.44 10.788 4.64
11 1-chloro-2,6-dimethyl-4-hydroxybenzene 9.549 9.476 −0.76 9.944 4.14
12 4-chloro-2-nitrophenol 6.48 7.724 19.20 6.475 −0.08
13 2-chlorophenol 8.55 8.974 4.96 8.117 −5.06
14 3-chlorophenol 9.10 8.72 −4.18 8.971 −1.42
15 4-chlorophenol 9.43 8.725 −7.48 9.03 −4.24
16 o-cresol 10.26 9.875 −3.75 10.174 −0.84
17 4-cyano-2,6-dimethylphenol 8.27 7.934 −4.06 7.872 −4.81
18 4-cyano-3,5-dimethylphenol 8.21 8.869 8.03 8.033 −2.16
19 3-cyanophenol 8.61 8.168 −5.13 8.2 −4.76
20 3,5-dibromophenol 8.056 7.186 −10.80 8.024 −0.40
21 2,4-dichlorophenol 7.85 8.141 3.71 7.96 1.40
22 2,6-dichlorophenol 6.78 7.264 7.14 6.827 0.69
23 3,5-diethoxyphenol 9.370 9.813 4.73 9.529 1.70
24 3-(diethoxyphosphinyl)phenol 8.68 9.267 6.76 8.628 −0.60
25 4-(diethoxyphosphinyl)phenol 8.28 8.517 2.86 8.276 −0.05
26 3,4-dihydroxybenzaldehyde 7.55 6.84 −9.40 7.623 0.97
27 1,2-dihydroxybenzene 9.356 10.487 12.09 9.456 1.07
28 1,4-dihydroxy-2,6-dinitrobenzene 4.42 2.184 −50.59 4.425 0.11
29 1,3-dihydroxy-2-methylbenzene 10.05 9.685 −3.63 9.603 −4.45
30 1,2-dihydroxy-3-nitrobenzene 6.68 3.728 − 44.19 6.68 0.00
31 1,2-dihydroxy-4-nitrobenzene 6.701 8.082 20.61 6.824 1.84
32 3,5-diiodophenol 8.103 6.653 −17.89 8.126 0.28
33 3,5-dimethoxyphenol 9.345 9.32 −0.27 9.497 1.63
34 2,6-dimethyl-4-nitrophenol 7.190 7.736 7.59 7.439 3.46
35 3,5-dimethyl-4-nitrophenol 8.245 8.131 −1.38 8.299 0.65
36 2,3-dimethylphenol 10.50 10.117 −3.65 10.246 −2.42
37 2,6-dimethylphenol 10.59 9.824 −7.23 10.264 −3.08
38 3,4-dimethylphenol 10.32 10.055 −2.57 10.298 −0.21
39 3,5-dimethylphenol 10.15 10.113 −0.36 10.068 −0.81
40 2,4-dinitrophenol 4.08 6.75 65.44 4.081 0.02
41 2,5-dinitrophenol 5.216 6.568 25.92 5.222 0.12
42 3,5-dinitrophenol 6.732 5.961 −11.45 6.658 −1.10
43 2-ethoxyphenol 10.109 10.886 7.69 10.117 0.08
44 3-ethoxyphenol 9.655 9.801 1.51 9.617 −0.39
45 2-ethylphenol 10.2 10.221 0.21 10.28 0.78
46 2-fluorophenol 8.73 9.247 5.92 9.112 4.38
47 3-fluorophenol 9.29 9.072 −2.35 9.16 −1.40
48 4-fluorophenol 9.89 9.078 −8.21 9.992 1.03
49 2'-hydroxyacetophenone 9.90 8.906 −10.04 9.232 −6.75
50 3'-hydroxyacetophenone 9.19 9.114 −0.83 9.46 2.94
51 3-hydroxybenzaldehyde 9.00 8.799 −2.23 9.25 2.78
52 4-hydroxybenzaldehyde 7.620 8.423 10.54 7.96 4.46
2012 Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 Aziz Habibi-Yangjeh et al.
Table 3. Continued

No. Compound Exp. MLR IPDMLR ANN IPDANN

53 2-hydroxybenzyl alcohol 9.92 8.355 −15.78 10.146 2.28
54 3-hydroxybenzyl alcohol 9.83 9.663 −1.70 9.866 0.37
55 1-hydroxy-2,4-dihydroxymethylbenzene 9.79 8.488 −13.30 9.615 −1.79
56 2-hydroxy-3-methoxybenzaldehyde 7.912 7.76 −1.92 7.974 0.78
57 (2-hydroxy-5-methylbenzene)-methanol 10.15 8.525 −16.01 9.876 −2.70
58 1-hydroxy-2-propylbenzene 10.50 10.439 −0.58 10.586 0.82
59 4-hydroxy-α,α,α-trifluorotoluene 8.675 8.982 3.54 8.559 −1.34
60 1-hydroxy-2,4,6-trihydroxymethylbenzene 9.56 8.357 −12.58 9.623 0.66
61 4-indanol 10.32 10.033 −2.78 10.289 −0.30
62 4-iodophenol 9.200 7.939 −13.71 9.088 −1.22
63 2,6-di-iodo-4-nitrophenol 3.32 4.131 24.43 3.304 −0.48
64 2-methoxyphenol 9.99 10.489 4.99 9.427 −5.64
65 2-methoxy-4-(2-propenyl)phenol 10.0 10.686 6.86 10.093 0.93
66 6-methyl-2-butylphenol 11.72 10.478 −10.60 11.065 −5.59
67 2-methyl-4-tert-butylphenol 10.59 10.963 3.52 11.007 3.94
68 2,2'-methylenebis(4-chlorophenol) 7.6 8.15 7.24 7.566 −0.45
69 2,2'-methylenebis(4,6-dichlorophenol) 5.6 6.129 9.45 5.637 0.66
70 4-methylsulfonyl-3,5-dimethylphenol 8.13 8.435 3.75 8.094 −0.44
71 3-(s-methylthio)phenol 9.53 8.572 −10.05 9.636 1.11
72 4-(s-methylthio)phenol 9.53 8.717 -8.53 9.743 2.24
73 2-nitrohydroquinone 7.63 8.45 10.75 7.633 0.04
74 2-nitrophenol 7.222 8.396 16.26 7.166 −0.78
75 4-nitrosophenol 6.48 7.768 19.88 6.693 3.29
76 phenol 9.99 9.578 −4.12 10.346 3.56
77 2-phenylphenol 9.55 8.847 −7.36 9.367 −1.92
78 5,6,7,8-tetrahydro-1-naphthol 10.28 9.883 −3.86 10.473 1.88
79 5,6,7,8-tetrahydro-2-naphthol 10.48 10.088 −3.74 10.639 1.52
80 2,4,6-tri-tert-butylphenol 12.19 11.724 −3.82 12.342 1.25
81 2,4,5-trichlorophenol 7.37 7.345 −0.34 7.396 0.35
82 3,4,5-trichlorophenol 7.839 7.275 −7.19 7.771 −0.87
83 3-trifluoromethylphenol 8.950 9.098 1.65 9.242 3.26
84 2,3,4-trimethylphenol 10.59 10.322 −2.53 10.648 0.55
85 2,4,5-trimethylphenol 10.57 10.325 −2.32 10.684 1.08
86 3,4,5-trimethylphenol 10.25 10.358 1.05 10.498 2.42
87 2,4,6-trimethylphenol 10.88 10.002 −8.07 10.58 −2.76
88 2,4,6-tripropylphenol 11.47 10.741 −6.36 11.165 −2.66
89 2-acetamidobenzoic acid 3.63 4.692 29.26 3.641 0.30
90 3-acetamidobenzoic acid 4.07 4.289 5.38 4.212 3.49
91 4-acetamidobenzoic acid 4.28 4.035 −5.72 3.87 −9.58
92 4-acetoxybenzoic acid 4.38 3.351 −23.49 3.979 −9.16
93 2-acetylbenzoic acid 4.13 4.025 −2.54 4.023 −2.59
94 3-acetylbenzoic acid 3.83 3.79 −1.04 3.597 −6.08
95 4-acetylbenzoic acid 3.70 3.654 −1.24 3.87 4.59
96 3-amino-1-naphthoic acid 2.61 3.817 46.25 2.834 8.58
97 anthracene-9-carboxylic acid 3.65 2.991 −18.05 3.562 −2.41
98 1,3-benzenedicarboxylic acid 3.62 3.374 −6.80 3.571 −1.35
99 1,4-benzenedicarboxylic acid 3.54 3.266 −7.74 3.975 12.29
100 1,2,4,5-benzenetetracarboxylic acid 1.92 2.563 33.49 2.002 4.27
101 1,2,3-benzenetricarboxylic acid 2.88 3.582 24.38 2.771 −3.78
102 1,2,4-benzenetricarboxylic acid 2.52 2.79 10.71 2.375 −5.75
103 1,3,5-benzenetricarboxylic acid 2.12 2.971 40.14 2.34 10.38
104 benzilic acid 3.09 3.872 25.31 3.27 5.83
105 benzylamine-4-carboxylic acid 3.59 4.358 21.39 4.14 15.32
106 2-biphenylcarboxylic acid 3.46 3.194 −7.69 3.665 5.92
107 2-bromobezoic acid 2.85 3.896 36.70 2.918 2.39
Prediction Acidity Constant Using QSPR Models Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 2013

Table 3. Continued

No. Compound Exp. MLR IPDMLR ANN IPDANN

108 3-bromobezoic acid 3.810 3.065 −19.55 3.915 2.76
109 3-tert-butylbenzoic acid 4.199 5.079 20.96 4.418 5.22
110 4-tert-butylbenzoic acid 4.389 4.985 13.58 4.104 −6.49
111 2-chlorobenzoic acid 2.877 3.287 14.25 3.086 7.26
112 3-chlorobenzoic acid 3.83 3.319 −13.34 3.698 −3.45
113 2-chloro-4-nitrobenzoic acid 1.96 1.619 −17.40 1.868 −4.69
114 2-chloro-5-nitrobenzoic acid 2.17 1.654 −23.78 1.929 −11.11
115 2-chloro-6-nitrobenzoic acid 1.342 2.153 60.43 1.417 5.59
116 2-cyanobenzoic acid 3.14 2.809 −10.54 2.962 −5.67
117 3,5-diaminobenzoic acid 5.30 4.731 −10.74 5.225 −1.42
118 3,6-dichlorophthalic acid 1.46 3.287 125.14 1.271 −12.95
119 2,4-dihydroxybenzoic acid 3.29 3.956 20.24 3.639 10.61
120 2,5-dihydroxybenzoic acid 2.97 3.802 28.01 3.408 14.75
121 3,5-dihydroxybenzoic acid 4.04 3.911 −3.19 3.805 −5.82
122 2,6-dimethoxybenzoic acid 3.44 5.312 54.42 3.447 0.20
123 2,3-dimethylbenzoic acid 3.771 4.55 20.66 3.668 −2.73
124 2,4-dimethylbenzoic acid 4.217 4.499 6.69 4.019 −4.70
125 2,5-dimethylbenzoic acid 3.990 4.522 13.33 3.799 −4.79
126 3,5-dimethylbenzoic acid 4.302 4.467 3.84 4.189 −2.63
127 2,3-dimethylnaphthalene-1-carboxylic acid 3.33 4.439 33.30 3.532 6.07
128 2,3-dinitrobenzoic acid 1.85 2.331 26.00 2.258 22.05
129 2,4-dinitrobenzoic acid 1.43 0.688 −51.89 1.587 10.98
130 2,5-dinitrobenzoic acid 1.62 1.12 −30.86 1.905 17.59
131 3,5-dinitrobenzoic acid 2.85 1.127 −60.46 2.681 −5.93
132 2-ethylbenzoic acid 3.79 4.575 20.71 3.757 −0.87
133 4-ethylbenzoic acid 4.35 4.451 2.32 4.2 −3.45
134 2-fluorobenzoic acid 3.27 3.629 10.98 3.67 12.23
135 3-fluorobenzoic acid 3.865 3.586 −7.22 3.584 −7.27
136 3-hydroxybenzoic acid 4.076 3.993 −2.04 4.496 10.30
137 4-hydroxybenzoic acid 4.582 3.912 −14.62 4.627 0.98
138 2-hydroxy-5-bromobenzoic acid 2.61 3.096 18.62 2.722 4.29
139 2-hydroxy-5-chlorobenzoic acid 2.63 3.342 27.07 3.394 29.05
140 4-hydroxy-3-methoxybenzoic acid 4.355 3.803 −12.68 4.022 −7.65
141 2-hydroxy-5-methylbenzoic acid 4.08 4.275 4.78 3.365 −17.52
142 2-hydroxy-6-methylbenzoic acid 3.32 5.268 58.67 3.231 −2.68
143 2-hydroxy-3-nitrobenzoic acid 1.87 2.427 29.79 1.93 3.21
144 2-hydroxy-5-nitrobenzoic acid 2.12 2.435 14.86 2.059 −2.88
145 2-hydroxy-6-nitrobenzoic acid 2.24 2.774 23.84 2.723 21.56
146 4-iodobenzoic acid 4.00 2.59 −35.25 4.162 4.05
147 mesitylenic acid 4.32 4.467 3.40 4.189 −3.03
148 2-methoxybenzoic acid 4.09 4.146 1.37 3.963 −3.11
149 3-methoxybenzoic acid 4.08 4.006 −1.81 4.305 5.51
150 3-methylbenzoic acid 4.269 4.282 0.30 4.303 0.80
151 4-methylbenzoic acid 4.362 4.214 −3.39 4.541 4.10
152 2-methyl-3,5-dinitrobenzoic acid 2.97 1.552 −47.74 2.982 0.40
153 2-methyl-1-naphthoic acid 3.11 3.509 12.83 3.135 0.80
154 3-methylsulfonylbenzoic acid 3.52 3.113 −11.56 3.581 1.73
155 4-methylsulfonylbenzoic acid 3.64 3.062 −15.88 3.323 −8.71
156 1-naphthalenecarboxylic acid 3.695 3.68 −0.41 3.938 6.58
157 2-naphthalenecarboxylic acid 4.161 3.214 −22.76 3.693 −11.25
158 4-nitrobenzene-1,2-dicarboxylic acid 2.11 1.933 −8.39 1.917 −9.15
159 2-nitrobenzoic acid 2.18 2.444 12.11 2.429 11.42
160 3-nitrobenzoic acid 3.46 2.38 −31.21 3.196 −7.63
161 4-nitrobenzoic acid 4.441 2.295 −48.32 3.286 −26.01
162 o-phthalic acid 2.950 3.361 13.93 2.555 −13.39
2014 Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 Aziz Habibi-Yangjeh et al.
Table 3. Continued

No. Compound Exp. MLR IPDMLR ANN IPDANN

163 3-sulfamylbenzoic acid 3.54 3.209 −9.35 3.73 5.37
164 4-sulfamylbenzoic acid 3.47 3.159 −8.96 3.324 −4.21
165 2,3,5,6-tetramethylbenzoic acid 3.415 5.746 68.26 3.401 −0.41
166 2,4,6-tribromobenzoic acid 1.41 2.063 46.31 1.408 −0.14
167 3,4,5-trihydroxybenzoic acid 4.19 3.528 −15.80 3.824 −8.74
168 2,4,6-trimethylbenzoic acid 3.448 5.211 51.13 3.641 5.60
Validation set
169 2-bromophenol 8.452 8.691 2.83 8.673 2.61
170 2,4-di-tert-butylphenol 11.64 11.741 0.87 11.887 2.12
171 4-chloro-2,6-dinitrophenol 2.97 1.531 −48.45 2.959 −0.37
172 m-cresol 10.00 9.835 −1.65 10.017 0.17
173 1,3-dichloro-2,5-dihydroxybenzene 7.30 7.295 −0.07 7.066 −3.21
174 3,4-dichlorophenol 8.630 8.014 −7.14 8.343 −3.33
175 1,3-dihydroxybenzene 9.44 9.497 0.60 9.415 −0.26
176 2,4-dimethylphenol 10.58 10.069 −4.83 10.367 −2.01
177 2,6-dinitrophenol 3.713 2.105 −43.31 3.718 0.13
178 3-ethylphenol 10.07 10.127 0.57 10.217 1.46
179 4'-hydroxyacetophenone 8.05 8.846 9.89 8.155 1.30
180 4-hydroxybenzyl alcohol 9.82 9.53 −2.95 9.829 0.09
181 3-hydroxy-4-methoxybenzaldehyde 8.889 6.87 −22.71 8.744 −1.63
182 2-iodophenol 8.464 8.042 −4.99 8.505 0.48
183 3-methoxyphenol 9.652 9.475 −1.83 9.698 0.48
184 3-methylsulfonylphenol 9.33 8.579 −8.05 9.219 −1.19
185 3-nitrophenol 8.360 7.627 −8.77 8.185 −2.09
186 4-phenylphenol 9.55 8.554 −10.43 9.647 1.02
187 1,2,3-trihydroxybenzene 9.03 8.412 −6.84 8.938 −1.02
188 2-acetoxybenzoic acid 3.48 3.934 13.05 3.743 7.56
189 4-amino-2-naphthoic acid 2.89 3.547 22.73 3.028 4.78
190 1,2,3,4-benzenetetracarboxylic acid 2.05 3.004 46.54 2.064 0.68
191 benzoic acid 4.204 4.015 −4.50 4.297 2.21
192 4-bromobezoic acid 3.99 2.879 −27.84 4.057 1.68
193 4-chlorobenzoic acid 3.986 3.234 −18.87 3.88 −2.66
194 3-cyanobenzoic acid 3.60 2.725 −24.31 3.381 −6.08
195 3,4-dihydroxybenzoic acid 4.48 3.758 −16.12 4.153 −7.30
196 2,6-dimethylbenzoic acid 3.362 4.656 38.49 3.486 3.69
197 2,6-dinitrobenzoic acid 1.14 2.051 79.91 1.103 −3.25
198 4-fluorobenzoic acid 4.14 3.573 −13.70 3.854 −6.91
199 2-hydroxy-3-methylbenzoic acid 2.99 4.336 45.02 3.376 12.91
200 2-iodobenzoic acid 2.86 3.547 24.02 2.922 2.17
201 4-methoxybenzoic acid 4.49 3.91 −12.92 4.456 −0.76
202 2-methyl-4-nitrobenzoic acid 1.86 2.746 47.63 2.968 59.57
203 2-nitrobenzene-1,4-dicarboxylic acid 1.73 1.981 14.51 1.953 12.89
204 2-phenoxybenzoic acid 3.53 3.36 −4.82 3.758 6.46
205 2,4,6-trihydroxybenzoic acid 1.68 1.793 6.73 1.793 6.73
Prediction set
206 3-bromophenol 9.031 8.378 −7.23 9.17 1.54
207 2,6-di-tert-butylphenol 11.7 11.053 −5.53 11.895 1.67
208 4-chloro-3-methylphenol 9.549 9.111 −4.59 9.593 0.46
209 p-cresol 10.26 9.774 −4.74 10.216 −0.43
210 2,3-dichlorophenol 7.44 8.196 10.16 7.827 5.20
211 3,5-dichlorophenol 8.179 7.873 −3.74 8.198 0.23
212 1,4-dihydroxybenzene 9.91 9.613 −3.00 10.153 2.45
213 1,4-dihydroxy-2,3,5,6-tetramethylbenzene 11.25 10.3 −8.44 10.723 −4.68
Prediction Acidity Constant Using QSPR Models Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 2015

Table 3. Continued
No. Compound Exp. MLR IPDMLR ANN IPDANN
214 2,5-dimethylphenol 10.22 10.115 −1.03 10.31 0.88
215 3,4-dinitrophenol 5.424 7.121 31.29 5.319 −1.94
216 4-ethylphenol 10.0 10.064 0.64 10.293 2.93
217 2-hydroxybenzaldehyde 8.34 9.833 17.90 8.155 −2.22
218 4-hydroxybenzonitrile 7.95 7.911 −0.49 8.166 2.72
219 4-hydroxy-3-methoxybenzaldehyde 7.396 7.974 7.82 7.896 6.76
220 3-iodophenol 8.879 8.099 −8.78 8.921 0.47
221 4-methoxyphenol 10.20 9.587 −6.01 10.282 0.80
222 4-methylsulfonylphenol 7.83 7.936 1.35 7.647 −2.34
223 4-nitrophenol 7.150 7.232 1.15 7.219 0.97
224 3-phenylphenol 9.63 8.485 −11.89 9.671 0.43
225 1,3,5-trihydroxybenzene 8.45 8.929 5.67 8.107 −4.06
226 3-acetoxybenzoic acid 4.00 3.47 −13.25 3.822 −4.45
227 anthracene-2-carboxylic acid 4.18 2.148 −48.61 4.186 0.14
228 1,2,3,5-benzenetetracarboxylic acid 2.38 1.625 −31.72 2.379 −0.04
229 2-benzoylbenzoic acid 3.54 3.223 −8.95 3.185 −10.03
230 2-bromo-6-nitrobenzoic acid 1.37 2.004 46.28 0.957 −30.15
231 2-chloro-3-nitrobenzoic acid 2.02 2.266 12.18 2.536 25.54
232 4-cyanobenzoic acid 3.55 2.619 −26.23 3.873 9.10
233 2,6-dihydroxybenzoic acid 1.30 2.864 120.31 1.084 −16.62
234 3,4-dimethylbenzoic acid 4.41 4.471 1.38 4.255 −3.51
235 3,4-dinitrobenzoic acid 2.82 2.251 −20.18 2.738 −2.91
236 2-hydroxybenzoic acid 2.98 4.091 37.28 3.313 11.17
237 2-hydroxy-4-methylbenzoic acid 3.17 4.308 35.90 3.128 −1.32
238 3-iodobenzoic acid 3.86 2.771 −28.21 3.529 −8.58
239 2-methylbenzoic acid 3.90 4.357 11.72 3.749 −3.87
240 2-methyl-6-nitrobenzoic acid 1.87 4.44 137.43 1.939 3.69
241 3-nitrobenzene-1,2-dicarboxylic acid 1.88 2.334 24.15 1.872 −0.43
242 4-phenoxybenzoic acid 4.52 3.194 −29.34 3.993 −11.66
a
Exp. refers to the experimental values of pKa, MLR and ANN refer to multi-parameter linear regression and artificial neural network calculated values
of pKa, respectively.

Figure 4 . Plot of the calculated values of pKa from the ANN model . Plot of the residual for calculated values of pKa from the
Figure 5

versus the experimental values of it for training, validation and ANN model versus the experimental values of it for prediction set.
prediction sets.
2016 Bull. Korean Chem. Soc. 2005, Vol. 26, No. 12 Aziz Habibi-Yangjeh et al.
Table 4. Comparsion of statistical parameters obtained by the MLR and ANN models for correlation acidity constant of phenols and
benzoic acids with the molecular descriptorsa
Model R2tot R2train R2valid R2pred RMSEtot RMSEtrain RMSEvalid RMSEpred
MLR 0.9266 0.9268 0.9400 0.9147 0.8610 0.8553 0.8034 0.9388
ANN 0.9931 0.9926 0.9943 0.9939 0.2648 0.2700 0.2479 0.2575
a
Subscript train is referring to the training set, valid is referring to the validation set and pred is referring to the prediction set, tot is refering to the total

data set and R is the correlation coefficient.

19. Zupan, J.; Gasteiger, J. Neural Networks in Chemistry and Drug

Conclusions Design; Wiley-VCH: Weinhein, 1999.
20. Agatonovic-Kustrin, S.; Beresford, R. J. Pharm. Biomed. Anal.
A linear and non-linear QSPR models have been develop- 2000 , 22, 717.
ed for prediction of acidity constant (pKa) for various 21. Fatemi, M. H. J. Chromatogr. A 2002 , 955, 273.
22. Xing, W. L.; He, X. W. Anal. Chim. Acta , 349, 283.
benzoic acids and phenols in water. Comparison of the
1997

23. Bunz, A. P.; Braun, B.; Janowsky, R. Fluid Phase Equilib. 1999 ,
values of RMSE for training, validation and prediction sets 158, 367.
(and other statistical parameters in Table 4) for the MLR and 24. Homer, J.; Generalis, S. C.; Robson, J. H. Phys. Chem. Chem.
ANN models show superiority of the nonlinear model over Phys. 1999 , 1, 4075.
the regression model. Root-mean square error of 0.9388 for 25. Goll, E. S.; Jurs, P. C. J. Chem. Inf. Comp. Sci. 1999 , 39, 974.
26. Vendrame, R.; Braga, R. S.; Takahata, Y.; Galvao, D. S. J. Chem.
the prediction set by the MLR model should be compared Inf. Comput. Sci. 1999 , 39, 1094.
with the value of 0.25751 for the ANN model. Since the 27. Gaspelin, M.; Tusar, L.; Smid-Korbar, J.; Zupan, J.; Kristl, J. Int.
improvement of the results obtained using nonlinear model J. Pharm. 2000 , 196, 37.
(ANN) is considerable, it can be concluded that the 28. Gini, G.; Cracium, M. V.; Konig, C.; Benfenati, E. J. Chem. Inf.
nonlinear characteristics of the molecular descriptors on the Comput. Sci. 2004 , 44, 1897.
29. Urata, S.; Takada, A.; Uchimaru, T.; Chandra, A. K.; Sekiya, A. J.
pKa values of the compounds in water is serious. Fluorine Chem. 2002 , 116, 163.
30. Koziol, J. Internet Electron J. Mol. Des. , 2, 315.
Acknowledgements. The Authors wish to acknowledge
2003

31. Wegner, J. K.; Zell, A. J. Chem. Inf. Comput. Sci. 2003 , 43, 1077.
the vice-presidency of research, university of Mohaghegh 32. Valkova, I.; Vracko, M.; Basak, S. C. Anal. Chim. Acta 2004 , 509,
Ardebili, for financial support of this work. 179.
33. Sebastiao, R. C. O.; Braga, J. P.; Yoshida, M. I. Thermochimica
Acta , 412, 107.
References 2004

34. Jalali-Heravi, M.; Masoum, S.; Shahbazikhah, P. J. Magn. Reson.

2004 , 171, 176.
1. Katritzky, A. R.; Karelson, M.; Lobanov, V. S. Pure Appl. Chem. 35. Habibi-Yangjeh, A.; Nooshyar, M. Bull. Korean Chem. Soc. 2005 ,
1997 , 69, 245. 26, 139.
2. Balaban, A. T. J. Chem. Inf. Comut. Sci. , 37, 645.
1997 36. Habibi-Yangjeh, A.; Nooshyar, M. Physics and Chemistry of
3. Benfenati, E.; Gini, G. Toxicology , 119, 213.
1997 Liquids 2005 , 43, 239.
4. Cronce, D. T.; Famini, G. R.; Soto, J. A. D.; Wilson, L. Y. J. 37. Selassie, C. D.; DeSoyza, T. V.; Rosario, M.; Gao, H.; Hansch, C.
Chem. Soc., Perkin Trans. 2 , 1293.
1998 Chemico-Biological Interaction 1998 , 113, 175.
5. Engberts, J. B. F. N.; Famini, G. R.; Perjessy, A.; Wilson, L. Y. J. 38. Zhao, Y.-H.; Yuan, L.-H.; Wang, L.-S. Bull. Environ. Contam.
Phys. Org. Chem. 1998, 11, 261. Toxicol. 1996 , 57, 242.
6. Hiob, R.; Karelson, M. J. Chem. Inf. Comut. Sci. , 40,
2000 39. Hemmateenejad, B.; Sharghi, H.; Akhond, M.; Shamsipur, M. J.
1062. Solution Chem. 2003, 32, 215.
7. Habibi-Yangjeh, A. Indian J. Chem. , 42B, 1478.
2003 40. Gruber, C.; Buss, V. Chemosphere 1989 , 19, 1595.
8. Habibi-Yangjeh, A. Indian J. Chem. 43B, 1504.
2004, 41. Citra, M. J. Chemosphere 1999 , 38, 191.
9. Nikolic, S.; Milicevic, A.; Trinajstic, N.; Juric, A. Molecules 42. Schuurmann, G. Quant. Struct. Act. Relat. 1996 , 15, 121.
2004 , 9, 1208. 43. Gross, K. C.; Seybold, P. G. Int. J. Quant. Chem. 2001 , 85, 569.
10. Devillers, J. SAR and QSAR Environ. Res. , 15, 501.
2004 44. Liptak, M. D.; Gross, K. C.; Seybold, P. G.; Feldgus, S.; Shields,
11. Karelson, M.; Lobanov, V. S. Chem. Rev. , 96, 1027.
1996 G. C. J. Am. Chem. Soc. 2002 , 124, 6421.
12. Todeschini, R.; Consonni, V. Handbook of Molecular Descriptors; 45. Hanai, T.; Koizumi, K.; Kinoshita, T. J. Liq. Chromatogr. Relat.
Wiley-VCH: Weinheim, Germany, 2000. Technol. 2000 , 23, 363.
13. Kramer, R. Chemometric Techniques for Quantitative Analysis; 46. Ma, Y.; Gross, K. C.; Hollingsworth, C. A.; Seybold, P. G.;
Marcel Dekker: New York, 1998. Murray, J. S. J. Mol. Model 2004 , 10, 235.
14. Wold, S.; Sjöström, M. Chemom. Intell. Lab. Syst. , 44, 3.
1998 47. HyperChem, Release 7.0 for Windows, Molecular Modeling
15. Barros, A. S.; Rutledge, D. N. Chemomet. Intell. Lab. Syst. ,
1998 System; Hypercube Inc.: 2002.
40, 65. 48. Todeschini, R.; Consonni, V.; Pavan, M. Dragon Software,
16. Garkani-Nejad, Z.; Karlovits, M.; Demuth, W.; Stimpfl, T.; Version 2.1; 2002.
Vycudilik, W.; Jalali-Heravi, M.; Varmuza, K. J. Chromatogr. A 49. Dean, J. A. Lange’s Handbook of Chemistry, 15th Ed.; McGraw-
2004 , 1028, 287. Hill, Inc.: 1999.
17. Patterson, D. W. Artificial Neural Networks: Theory and 50. Demuth, H.; Beale, M. Neural Network Toolbox; Mathworks:
Applications; Simon and Schuster: New York, 1996; Part III, Natick, MA, 2000.
Ch. 6. 51. Despagne, F.; Massart, D. L. Analyst 1998 , 123, 157R.
18. Bose, N. K.; Liang, P. Neural Network Fundamentals; McGraw- 52. Matlab 6.5; Mathworks: 1984-2002.
Hill: New York, 1996. 53. Famini, G. R.; Wilson, L. Y. J. Phys. Org. Chem. 1999 , 12, 645.