INTRODUCTION:
Being used to a great degree as industrial chemicals, Aromatic amines—anilines and their related derivate compounds—represent a significant portion of environmental pollutants. Aniline is the molecule from which an immense set of aromatic amines find their origin. It has occupied a place in the top hundred most important building blocks of chemistry since its discovery in 1826. In a plethora of varied applications, Aniline and its derivatives—comprising of chloro substituents—are utilized in the form of intermediates.
In June 2007, Registration, Evaluation, Authorization and Restriction of Chemicals (REACH) legislation was set in motion1. This has led to a tremendous wave of interest in the field of silico methodologies. By inaugurating and enforcing the concept of ‘no data, no market’, REACH made certain that manufacturing and importation abided by a high standard of safety for both human health and the environment. With turning away from non-essential testing (notably animal testing) being a goal that is stated explicitly by REACH, the latter offered the parties involved a plethora of options that could potentially allow them to seek out the accomplishment of such objective. This includes alternative test methods like in vitro and in silico methodologies.
Quantitative structure-activity relationships (QSAR) analysis emerged as one of the aforementioned methodologies. Due to its low cost and time requirements, it has proven itself extremely valuable2–11.
What constitutes QSAR analysis is a miscellaneous set of mathematical and statistical procedures that are characterized by the purpose of discovering functional relationships between chemical compound structure—as depicted through experimental or theoretical variables labelled molecular descriptors 12—, and their measured properties and activities.
To develop a model that can simulate the relationship between the structures of 48 aromatic amines and their theoretical descriptors, we have, in our work, conducted a QSAR study. And in order to pinpoint and select the most informative descriptors out of the entirety of the descriptors calculated by E-calc (Version 5.4) software, we used the genetic algorithm. By relying on MLR and ANN (artificial neutral networks) as different approaches, we developed a model through the selected descriptors. The usefulness of such model lies in predicting the log (1/CIC50) (decimal logarithm of the inverts of the 50%inhibitory growth concentration (IGC50) of Tetrahymenapyriformis. In order to make sure the models are valid, we have divided the data set into a training set (35 compounds) and a test set (13 compounds), using the Kennard and Stones algorithm. Furthermore, to emphasize the structural prerequisites for an ideal aquatic toxicity, various statistical methods were core to the model’s development. Our research purpose can be divided into three major aims: First, the exploration of the relationships between structure and activity when it comes to the aquatic toxicity of compounds; second, the selection of the best possible predictive model among other similar chemometric models of aquatic toxicity, and third, using the two approaches MLR and ANN to test the level of performance and the stability of the obtained model.
METHODS:
The anilines and their derivate compounds were subject to a QSPR study. It was conducted thanks to the assistance of a data set selection, molecular descriptors generation, Multiple linear regression statistical analysis, as well as model validation techniques. Moreover, to further test how applicable the model is, the predicted data was plotted against the experimental data of compounds. The TSAR simulation software was the tool we relied on to perform all regression analysis. The following statistical parameters served as a testing tool for predicting the efficiency of the obtained model: correlation coefficient (r), coefficient of determination (r2), standard deviation (s), and null hypothesis test (F-test). The cross-validation coefficient (rcv2) was crucial for testing the model’s validity via relying on the “leave-one-out” feature in the software Moby Digs for Windows13. Otherapproaches such as Artificial neural networks14affirmed stability and prediction accuracy.
Experimental Data:
The recognized experimental values of pIGC50 (Table I) represent a distinctly dissimilar data set that is made up of 48 aromatic amines compounds. It was taken from literature on the subject matter15.
Calculation of descriptors:
Through HYPERCHEM Software(2002), each compound’s chemical structure is outlined in a computer 16, and then, it is pre-optimized via the MM+ molecular mechanics method (Polak-Ribiere algorithm). In order to find the conclusive geometries of minimum energy conformation, the PM3 method—semi-empirical—was used at a restricted Hartree-Fock level, plus no configuration interaction, and while a gradient norm limit of 0.001 kcal Å−1 mol−1 is applied as a stopping criterion. After that, the geometries were used as input for the generation of 1664 descriptors via Dragon software (Version 5.4)17.
Kennard and Stones algorithm:
One of the most commonly used algorithms for separating a given data sets into two distinguishable subsets is the Kennard and Stones algorithm. In essence, at first it attempts to detect two samples on the basis of the farthermost input variables from one another. The samples are then taken out from the original set of data, and placed into a calibration set. Rinse and repeat until the sample number selected in the calibration set proves satisfactory. The scope of these calibration samples is always within the measured region of the overall input variable space. It also is compatible with the induced metric, and the no validation samples are always placed outside the score of the measured region. These are the main advantages of this particular algorithm, and that is why it is deemed as one of the most important methods when it comes to build training and test sets. 18,19
Descriptor selection:
By relying on the Ordinary Least Square Regression method and the GA-variable subset selection 13, we conducted both the MLR analysis and variable selection via Todeschini et al’ Moby Digs software. GAs result in a population of one hundred regression models that are organized in accordance with decreasing internal predictive performance. As they are examined through Q2, the lower the latter is, the fewer the descriptors.
a) Electrotopological descriptor:
Kier and Hall introduced an entirely new paradigm, and it was reexamined lately. This paradigm is called the electrotopological state. It is meant to represent both the electronic and topological properties of the atom12. Together with the perturbation exercised by other atoms in the same molecule, the intrinsic value of an atom—in its valence state—represent its E-state. The Kier-Hall electronegativity model constitutes the basis for an atom’s intrinsic state. It is deduced from the electro negativity to skeletal bonds o ratio.
(1)
and 
are the molecularconnectivityvalues. They are given as follows (for
first row atoms)
(2)
With = 
-, h = skeleton connections number(2)
whereis electrons number in T orbitals; h is
the number of atom-bonded hydrogenatoms.
represents valence electrons number; 
, on the other hand, is the number of
electrons in orbitals; 
is the lone pair number of electrons; N
is the principal quantum number of the valence shell for that atom. We
can say that- isequal to the numberofand lone pair electrons. It was shown by
Kier and Hall to be proportional to the valence state electronegativity
b) Estimation of Octanol/Water partition coefficient
AlogP (Atomic increment system-based Ghose and Cripen model)19
When it comes to fragments being defined on an exclusively atomic level, several models dealing with that can be found in the literature. The advantage of such models lies in the reduced complexity of both fragment recognition and calculation—correction substructures unapplied (see Eq. (1)). As for the most common atomic increment system, AlogP, it originates in the works of Ghose and Crippen20. In this system, the neighboring environment of atoms is what classifies them. For carbon atoms, it is also their hybridization. The estimated logP for a given compound is as follows:
Alogp = ∑ni ai (3)
ni stands for ith atom type occurrence, and ai is the hydrophobicity constant. In the DRAGON software, the AlogP model at hand was assessed through a group of 2648 compounds—known experimental logP come from the NCI open data base—the resulting correlation coefficient r was equal to 0.915.
Chemometric Methods. (ANN-MLR):
a)MLR:
MLR can be defined as a statistical tool that functions by regressing independent variables against a dependent variable. MLR’s purpose is determining the property of interest’s linear model. It is expressed as:
y = a0 +𝛴ai xi (4)
y stands for the property, i.e. the dependent variable; xi, on the other hand, represents the molecular descriptors; and ai descriptor coefficient, while a0 is the symbol for the equation intercept.
b)ANN:
Is it possible that the human brain and its functions are being simulated and modelled by artificial systems? Such a question lies at the core principle of ANN. The latter is an approach that is comprised of the following elements: Processing elements, i.e. nodes, connection topology between nodes, and the acquisition of the principle that govern the encoding of new information into the network. There are certainly several models of ANN, nevertheless, the most used in QSAR is represented by the three-layered feed-forward network14. In this particular type, neurons are placed into an input, a hidden layer, and an output layer. Any given neuron in any given layer is characterized by a connection with the neuron of the layer that comes after it. On the other hand, connection between same-layer neurons is inexistent.
Model development and validation:
We used the features of Ordinary Least Squares Regression (OLS) and GA-VSS (GA for variable subset selection) in the software packing Moby Digs for Windows 13 to conduct both the multiple linear regression analysis and the variable selection. The extent to which the model is functioning was tested via the multiple determination coefficient, the R2, as well as the standard deviation error in calculation (SDEC). It is penned as follows:
(5)
Assessing internal prediction levels (Q2LMOcross-validation, bootstrap) and model robustness model (Q2LOO cross-validation, Y-scrambling) can be achieved through cross validation techniques.
Cross validation via the LOO (leave-one-out) technique relies on an n training set of n-1 objects that are included in the test set, and on the process of predicting the excluded ones. The cross validated explained Q2LOO is defined as:
=1- (6)
With y I and 
representing
respectively the measured and averaged (over the entire data set) values of the
dependent variable ; y ii denotes the response of the i-th
estimated by using a model without the i-th object; the summations
run over all compounds in the training set. The dispersion of predicted values
is measured via the PRESS (predictive residual sum of squares). It defines Q2,
and SDEP.
(7)
A value Q2> 0.5 is generally regarded as a good result and Q2> 0.9 as excellent 21.
Several works, however, proved that despite Q2 being essential for a high level of prediction in a model, it still is not enough.
When it comes to the method of bootstrap validation K, the generation of n-dimensional groups is conducted via a selection of n-objects—repeated randomly—from original data set. The first selected objects provide a first model that allows for the prediction of excluded sample. Only then is Q2 is measured for each and every model. The process of bootstrapping is conducted for 8000 times for each valid model. Through the chosen model 22, test objects response values are calculated; prediction quality is determined in terms of Q2ext23:
(8)
next being a representative of the total of objects belonging to the set left out by bootstrap, i.e. the external set. While ntr stands for the number of training set objects.
R² is also a very useful factor for validating chemicals via implementing the resulting model on the training set. An equally important one is the SPEDext (external standard deviation error of prediction) and it is penned as follows:
(9)
The sum summarizes the total of the set objects of the test (n)ext.
The following equation determines the external R²CVext for the test set:
(10)
and 
constitute,
respectively, the observed and calculated response values. While 
tra
represents the training set’s averaged value for the response
variable; the total sums up the test set.
QSAR AD (Applicability Domain):
The Williams plot of jackknife residuals vs the leverages, i.e. diagonal values (hi), was utilized to examine the Applicability Domain. Jackknife residuals—studentized residuals—can be defined as standardized and cross-validated. Each residual’s standard deviation is calculated with no reference to the i-th observation. Afterwards, each residual is divided by its corresponding SD. The leverage (hi) is the chemical’s value in the original variable space; it is as follows:
hi= xi (X T X)-1xiT (i=1,…..,n) (11)
The query compound’s descriptor row-vector is expressed by xi. For n training set compounds, the n (p+1) matrix of p model parameter values is represented by X. The matrix/vector’s transpose is denoted by the superscript T, while (3p + 1)/n is what defines the warning leverage value of (h*)24. Furthermore, when the value h of a compound is lower than h *, the predicted and actual values are characterized by a likelihood of compatibility equal to that of training set compounds. For instance, a given chemical with an h i> h* might support the model, if part of the training set. Nevertheless, it still remains unreliable when it comes to predicted data and the associated validation set. It could, in addition, not seem as an outlier due to the residual related to it being low. That is the reason why for a reliable portrayal of the AD, the leverage and the jackknife residual ought to be combined.
Permutation Test:
For the purpose of evaluating how reliable is the model training process, we conducted a randomization test. Selecting features and building predictive models were both conducted through the use of 75 randomly-labeled samples. As a deciding index, we used the models’ training set prediction accuracy, so that the level of performance of both the integrated model and the Y-randomization model can be compared to the five leading molecular descriptors and a signature gene 2526.
RESULTS AND DISCUSSION:
Applying the GA-VSS resulted in various reliable models related to the prediction that is based on several sets of molecular descriptors. The best possible model came as a result of using the electrotopological descriptor. In Table 1, the partition coefficient data of descriptor value and aquatic toxicity are listed.