336
A. Zakharov and A. Lagunin
logical effects of some molecules in the biological system, e.g. the ligand-receptor
interaction has a distinct isomeric specificity. Influence of three-dimensional characteristics of ligand-receptor interactions, such as dimensional alignment of molecules and fitting electronic properties of molecules based on their surfaces can be
very significant in pharmacology [26]. For example, the interaction of ligands with
CYP2C9 was extended to the fourth dimension (conformer’s analysis) and led to a
4-D classification of drugs [27]. The multi-dimensional relationships between precursors of anabolic steroids and mineral corticoid receptor were modeled and it was
shown how these substances disrupted the endocrine system [28].
11.2.4 Mathematical Methods
There are several handbooks [29, 30, 31] with descriptions of mathematical methods (machine learning techniques) used in QSAR modeling. The most known methods are: Naïve Bayes, Decision Trees, Fuzzy Logic, Genetic Algorithms, Multiple
Regressions, Neural Networks, Partial Least Squares, Radial Basis Function, Support Vector Machines. Here, we describe the most well-known computer programs,
providing the algorithms, used to build (Q)SAR models (Table 11.3).
Currently, the main tendency in the development of QSAR modeling is the use
of consensus models. When several models with different machine learning techniques and/or different sets of descriptors are developed based on the same training
set. The consensus model results in aggregation of predictions from all developed
models. Predictions from the models can be arithmetically averaged (simple unweighted consensus) or can be averaged with some weights for each model (weighted consensus) [32, 33]. It is considered that the use of the consensus model reduces
the variability of the individual models, which leads to more reliable predictions
[34]. These statements are valid for both QSAR, and SAR models.
11.3 The Practical Use of the Methods for Computational
Toxicity Prediction
Toxicological (Q)SAR models are essentially used for the toxicity prediction of
new compounds. These models are developed mainly from the training set of compounds with a known activity. If the training set is large and diverse (chemically
heterogeneous), then (Q)SAR models based on this set are considered to be global.
If the training set consists of a homogeneous compound, the (Q)SAR models, based
on this set, are called as local. The assessment of existing (Q)SAR methods to predict most significant toxicological values was performed during preparation of the
computational toxicology report by the European Commission under REACH development (early 2000’s). According to this report the accurate (Q)SAR models
were developed based on non-heterogeneous data (local (Q)SAR models). Nowadays the situation has been considerably changed. At the present time a reasonable
A. Zakharov and A. Lagunin
logical effects of some molecules in the biological system, e.g. the ligand-receptor
interaction has a distinct isomeric specificity. Influence of three-dimensional characteristics of ligand-receptor interactions, such as dimensional alignment of molecules and fitting electronic properties of molecules based on their surfaces can be
very significant in pharmacology [26]. For example, the interaction of ligands with
CYP2C9 was extended to the fourth dimension (conformer’s analysis) and led to a
4-D classification of drugs [27]. The multi-dimensional relationships between precursors of anabolic steroids and mineral corticoid receptor were modeled and it was
shown how these substances disrupted the endocrine system [28].
11.2.4 Mathematical Methods
There are several handbooks [29, 30, 31] with descriptions of mathematical methods (machine learning techniques) used in QSAR modeling. The most known methods are: Naïve Bayes, Decision Trees, Fuzzy Logic, Genetic Algorithms, Multiple
Regressions, Neural Networks, Partial Least Squares, Radial Basis Function, Support Vector Machines. Here, we describe the most well-known computer programs,
providing the algorithms, used to build (Q)SAR models (Table 11.3).
Currently, the main tendency in the development of QSAR modeling is the use
of consensus models. When several models with different machine learning techniques and/or different sets of descriptors are developed based on the same training
set. The consensus model results in aggregation of predictions from all developed
models. Predictions from the models can be arithmetically averaged (simple unweighted consensus) or can be averaged with some weights for each model (weighted consensus) [32, 33]. It is considered that the use of the consensus model reduces
the variability of the individual models, which leads to more reliable predictions
[34]. These statements are valid for both QSAR, and SAR models.
11.3 The Practical Use of the Methods for Computational
Toxicity Prediction
Toxicological (Q)SAR models are essentially used for the toxicity prediction of
new compounds. These models are developed mainly from the training set of compounds with a known activity. If the training set is large and diverse (chemically
heterogeneous), then (Q)SAR models based on this set are considered to be global.
If the training set consists of a homogeneous compound, the (Q)SAR models, based
on this set, are called as local. The assessment of existing (Q)SAR methods to predict most significant toxicological values was performed during preparation of the
computational toxicology report by the European Commission under REACH development (early 2000’s). According to this report the accurate (Q)SAR models
were developed based on non-heterogeneous data (local (Q)SAR models). Nowadays the situation has been considerably changed. At the present time a reasonable