467
14 Quantitative Structure-Property Relationship Analysis …
predicting API bioavailability. Realization of such project requires information on
physiological nature of bioavailability and physico-chemical properties of API.
Now the market of software for prediction of biological properties of chemical
substances is saturated with various products. Analysis of this problem has been
reviewed in [21–23].
For significant prediction of bioavailability a training set of drugs is subdivided
into either two (e.g. Yes/No, Positive/Negative), three [24] (high with F > 80 %;
moderate with F = 21–79 %, and low with F < 20 %), or four groups [25] (1: < 20 %,
2: 20–49 %; 3: 50–79 %; 4: > 80 %).
While experimental methods always require sufficient amount of chemicals for
the estimation of drug absorption, computational ( in silico) methods can lead to
the prediction of intestinal absorption based on chemical structure, and can thus be
used before synthesis of compounds. In silico predictions could be based both on
relatively simple quantitative structure-activity relationships (QSAR) analysis and
more complex physiologically based pharmacokinetic and/or pharmacodynamic
models. Whichever the approach used for model building, computational methods
should be based on experimental data that were obtained for a wide range of structurally diverse compounds (training set). It should be noted, however, that current
in silico methods, are not as reliable as experimental models.
The models can be divided into two categories: regression and classification
ones. Most of the published studies are based on small or large sets of permeability
data collected from literature sources. Since the validity of any model primarily
depends on the data, a large and accurate dataset is required for the development
of global models with general applicability. Model validation is another critically
important step in building robust QSPR models. Ideally, one should report results
based on the training set, the cross-validation set, and an external dataset to increase
the user’s confidence level. Consistent reporting of model statistics is highly desirable so that readers can objectively evaluate the model quality and applicability in
a real-life drug discovery setting [22].
Rather than trying to predict specific absorption-related quantities, researchers
have tried to find general principles to distinguish drug-like from non-drug-like
molecules by analyzing databases of drugs and non-drugs. Generally, these rules
obtained from database analysis can be used to distinguish well-absorbed molecules
from poorly-absorbed molecules. Among numerous attempts to find relevant correlation between physico-chemical properties of API and their bioavailability the
“rule of five” is the most popular ADME-concerned filters, and most widely used
[24]. The “rule of five” defined several rules for identifying compounds with possible poor absorption and permeability: (1) molecular weight > 500, (2) calculated
log P > 5 (CLOGP) or > 4.15 (MLOGP), (3) number of hydrogen bond donors (OH
and NH groups) > 5, and (4) number of hydrogen-bond acceptors (N and O atoms) > 10. A disadvantage of the “rule of five” is that it can only give a quite rough
classification of molecules, allowing the elimination of only a very limited set of
molecules. Later Lipinski [26] indicated that this rule does not belong to ideal filters, because predictions based on these rules are basically related to one parameter
Mlog P or Clog P, where errors are possible during additive evaluation (CLOGP
14 Quantitative Structure-Property Relationship Analysis …
predicting API bioavailability. Realization of such project requires information on
physiological nature of bioavailability and physico-chemical properties of API.
Now the market of software for prediction of biological properties of chemical
substances is saturated with various products. Analysis of this problem has been
reviewed in [21–23].
For significant prediction of bioavailability a training set of drugs is subdivided
into either two (e.g. Yes/No, Positive/Negative), three [24] (high with F > 80 %;
moderate with F = 21–79 %, and low with F < 20 %), or four groups [25] (1: < 20 %,
2: 20–49 %; 3: 50–79 %; 4: > 80 %).
While experimental methods always require sufficient amount of chemicals for
the estimation of drug absorption, computational ( in silico) methods can lead to
the prediction of intestinal absorption based on chemical structure, and can thus be
used before synthesis of compounds. In silico predictions could be based both on
relatively simple quantitative structure-activity relationships (QSAR) analysis and
more complex physiologically based pharmacokinetic and/or pharmacodynamic
models. Whichever the approach used for model building, computational methods
should be based on experimental data that were obtained for a wide range of structurally diverse compounds (training set). It should be noted, however, that current
in silico methods, are not as reliable as experimental models.
The models can be divided into two categories: regression and classification
ones. Most of the published studies are based on small or large sets of permeability
data collected from literature sources. Since the validity of any model primarily
depends on the data, a large and accurate dataset is required for the development
of global models with general applicability. Model validation is another critically
important step in building robust QSPR models. Ideally, one should report results
based on the training set, the cross-validation set, and an external dataset to increase
the user’s confidence level. Consistent reporting of model statistics is highly desirable so that readers can objectively evaluate the model quality and applicability in
a real-life drug discovery setting [22].
Rather than trying to predict specific absorption-related quantities, researchers
have tried to find general principles to distinguish drug-like from non-drug-like
molecules by analyzing databases of drugs and non-drugs. Generally, these rules
obtained from database analysis can be used to distinguish well-absorbed molecules
from poorly-absorbed molecules. Among numerous attempts to find relevant correlation between physico-chemical properties of API and their bioavailability the
“rule of five” is the most popular ADME-concerned filters, and most widely used
[24]. The “rule of five” defined several rules for identifying compounds with possible poor absorption and permeability: (1) molecular weight > 500, (2) calculated
log P > 5 (CLOGP) or > 4.15 (MLOGP), (3) number of hydrogen bond donors (OH
and NH groups) > 5, and (4) number of hydrogen-bond acceptors (N and O atoms) > 10. A disadvantage of the “rule of five” is that it can only give a quite rough
classification of molecules, allowing the elimination of only a very limited set of
molecules. Later Lipinski [26] indicated that this rule does not belong to ideal filters, because predictions based on these rules are basically related to one parameter
Mlog P or Clog P, where errors are possible during additive evaluation (CLOGP