328
A. Zakharov and A. Lagunin
3. A defined domain of applicability. It is a necessary to estimate an applicability
domain for each developed model;
4. Appropriate measures of goodness for fit, robustness and predictivity. It is a necessary to validate the developed models according to different types of validation
procedure;
5. A mechanistic interpretation, if possible. Each model, if possible, should have a
mechanistic interpretation;
The analysis of key components in QSAR modeling procedure, which includes
common mistakes, is given below.
11.2.1 Data
A considerable attention should be given to data, which are used for the models
construction. The following criteria are used for this purpose:
1. Reliability: accurate and complete description of the experimental values;
Data should be obtained by the same experimental protocol (using the same type
and sex of the animal, route of administration, time of exposure);
For quantitative parameters (LD 50 , EC 50 , etc.) the molar concentrations (mmol/
kg) should be used instead of weight (mg/kg) or volume (ppm). The following
equations are used for conversion of weight and volume values to the molar
concentrations:
mMole kg
mg kg
g mole
mMole m
ppm
/
/
/
,
/
.
=
=
3
24 25
The structure of compounds should be a single-component and presented in a
neutral form;
References or literature sources should be provided for each type of data;
2. Consistency: experimental results must be reproducible with low error;
For qualitative data (for example, a carcinogen and not a carcinogen) should not
be contradictions;
Classification of the compounds should be made according to the same criteria;
3. Reproducibility: experimental results should be reproducible in different
laboratories.
In addition, it is a necessary to take into account the range of a dependent variable
during the construction of (Q)SAR models. Gedeck et al. has showed on different
datasets that the prediction accuracy is significantly increased with magnification
of a value range in the dependent variable [12]. The authors have proved that to develop a good model, it is required to use the data, which has the range of dependent
values as at least one logarithm (the value of Y (e.g., LD 50 ) varies by 10 times).
A. Zakharov and A. Lagunin
3. A defined domain of applicability. It is a necessary to estimate an applicability
domain for each developed model;
4. Appropriate measures of goodness for fit, robustness and predictivity. It is a necessary to validate the developed models according to different types of validation
procedure;
5. A mechanistic interpretation, if possible. Each model, if possible, should have a
mechanistic interpretation;
The analysis of key components in QSAR modeling procedure, which includes
common mistakes, is given below.
11.2.1 Data
A considerable attention should be given to data, which are used for the models
construction. The following criteria are used for this purpose:
1. Reliability: accurate and complete description of the experimental values;
Data should be obtained by the same experimental protocol (using the same type
and sex of the animal, route of administration, time of exposure);
For quantitative parameters (LD 50 , EC 50 , etc.) the molar concentrations (mmol/
kg) should be used instead of weight (mg/kg) or volume (ppm). The following
equations are used for conversion of weight and volume values to the molar
concentrations:
mMole kg
mg kg
g mole
mMole m
ppm
/
/
/
,
/
.
=
=
3
24 25
The structure of compounds should be a single-component and presented in a
neutral form;
References or literature sources should be provided for each type of data;
2. Consistency: experimental results must be reproducible with low error;
For qualitative data (for example, a carcinogen and not a carcinogen) should not
be contradictions;
Classification of the compounds should be made according to the same criteria;
3. Reproducibility: experimental results should be reproducible in different
laboratories.
In addition, it is a necessary to take into account the range of a dependent variable
during the construction of (Q)SAR models. Gedeck et al. has showed on different
datasets that the prediction accuracy is significantly increased with magnification
of a value range in the dependent variable [12]. The authors have proved that to develop a good model, it is required to use the data, which has the range of dependent
values as at least one logarithm (the value of Y (e.g., LD 50 ) varies by 10 times).