374
P. M. Vassiliev et al.
activity level. For example, activity may be described as “absent-present,” “highlow” (in the opinion of an expert pharmacologist), or “fitting-missing” (when we refer to a range of some quantitative evaluation, such as 100 < LD 50 < 500 mg/kg). For
a quantitative evaluation, this prediction is carried out according to several binary
oppositions describing the activity; for example, “high-other,” “moderate-other,” or
“low-other.” To enhance the adequacy and consistency of the prediction, associated
classes of activity are formed, such as “high or moderate-other”, “high or moderate or low-other” (“active-inactive”). This delineation usually suffices for practical
purposes because the experimenting pharmacologist is typically interested in highly
active compounds, which he or she will test on a first-priority basis.
In IT Microcosm, prediction is made on 11 levels by describing the chemical
structure by various types of descriptors using four mathematical methods; the
spectrum of intermediate prediction estimates is generalized on the basis of three
voting strategies, and the prediction results are generalized for all strategies, while
the spectrum of prediction estimates is simultaneously checked for noncontradiction [96, 102, 103, 105, 110, 127].
In a training set, the classification dependences are established by methods of
pattern recognition and machine learning. These regularities unite various activity
levels that are pre-set in the form of semiquantitative gradations, and the structure
of a given compound is represented as a matrix of structural descriptors. The chemical structure is represented by 11 types of descriptors in a specialized hierarchic
multilevel language, QL [107, 110]. These descriptors form generalized patterns of
classes of active/inactive compounds, represented as matrices of structural descriptors [109] within the framework of a generalized pattern of a compound class with
the desired properties [94, 101, 105].
When predicting the presence or extent of a desired pharmacological activity, the
structural formula of an untested compound represented as a standard connection
table is transformed by the translator program into descriptors in the working language. By comparing the obtained pattern with models of class patterns using four
prediction methods that are essentially different in their mathematical formalism, a
spectrum of 44 intermediate prediction estimates (11 for each type of QL descriptor)
is calculated for each type of activity. This spectrum is then generalized on the basis
of one of three strategies, and a final estimate of the predicted compound activity is
calculated [103, 105, 109].
At the final stage, the prediction results are generalized in relation to all strategies, and the spectrum of prediction estimates is checked for noncontradiction
[102]. To enhance the reliability of this method, one can consolidate the prediction
results in relation to several levels of a predicted pharmacological activity [133].
The integral decision rules that are generated by IT Microcosm are consensus
QSAR regularities of the fourth level; the first consensus level relates to the 11
types of QL descriptors, the second level relates to the four prediction methods, the
third level consists of the three prediction strategies, and the fourth level is based on
the levels of the predicted activity [134].
P. M. Vassiliev et al.
activity level. For example, activity may be described as “absent-present,” “highlow” (in the opinion of an expert pharmacologist), or “fitting-missing” (when we refer to a range of some quantitative evaluation, such as 100 < LD 50 < 500 mg/kg). For
a quantitative evaluation, this prediction is carried out according to several binary
oppositions describing the activity; for example, “high-other,” “moderate-other,” or
“low-other.” To enhance the adequacy and consistency of the prediction, associated
classes of activity are formed, such as “high or moderate-other”, “high or moderate or low-other” (“active-inactive”). This delineation usually suffices for practical
purposes because the experimenting pharmacologist is typically interested in highly
active compounds, which he or she will test on a first-priority basis.
In IT Microcosm, prediction is made on 11 levels by describing the chemical
structure by various types of descriptors using four mathematical methods; the
spectrum of intermediate prediction estimates is generalized on the basis of three
voting strategies, and the prediction results are generalized for all strategies, while
the spectrum of prediction estimates is simultaneously checked for noncontradiction [96, 102, 103, 105, 110, 127].
In a training set, the classification dependences are established by methods of
pattern recognition and machine learning. These regularities unite various activity
levels that are pre-set in the form of semiquantitative gradations, and the structure
of a given compound is represented as a matrix of structural descriptors. The chemical structure is represented by 11 types of descriptors in a specialized hierarchic
multilevel language, QL [107, 110]. These descriptors form generalized patterns of
classes of active/inactive compounds, represented as matrices of structural descriptors [109] within the framework of a generalized pattern of a compound class with
the desired properties [94, 101, 105].
When predicting the presence or extent of a desired pharmacological activity, the
structural formula of an untested compound represented as a standard connection
table is transformed by the translator program into descriptors in the working language. By comparing the obtained pattern with models of class patterns using four
prediction methods that are essentially different in their mathematical formalism, a
spectrum of 44 intermediate prediction estimates (11 for each type of QL descriptor)
is calculated for each type of activity. This spectrum is then generalized on the basis
of one of three strategies, and a final estimate of the predicted compound activity is
calculated [103, 105, 109].
At the final stage, the prediction results are generalized in relation to all strategies, and the spectrum of prediction estimates is checked for noncontradiction
[102]. To enhance the reliability of this method, one can consolidate the prediction
results in relation to several levels of a predicted pharmacological activity [133].
The integral decision rules that are generated by IT Microcosm are consensus
QSAR regularities of the fourth level; the first consensus level relates to the 11
types of QL descriptors, the second level relates to the four prediction methods, the
third level consists of the three prediction strategies, and the fourth level is based on
the levels of the predicted activity [134].