378
P. M. Vassiliev et al.
2. No reduction of the dimensionality of mega-dimensional space of description
is carried out; the totality of the parameters is assumed to be informative. As a
consequence, no information determining the specifics of the objects to be recognized is lost, which allows for a more effective extrapolation of the obtained
regularities to new or poorly studied compounds with nontrivial specifics of
action.
3. A higher adequacy and enhanced prognostic capacity of the decision rules is
achieved by expanding the feature space by adding new parameter groups and
new levels of compound structure representation, rather than selecting “significant” variables or reducing the feature space.
4. This approach develops an ensemble of decision rules that are based on several
classification methods that differ essentially in their mathematical formalism
instead of a “better” prediction regularity. An independent classification of the
properties of the predicted object is done by each method for each description
level of each parameter group. The obtained spectrum of prediction estimates of
chemical compound activity is then used in final classification procedures.
5. The context-independence of the integral decision rule is a result of the generalization of the spectrum of intermediate prediction estimates of activity by the
methods of decision-making theory. As a result, primary decision rules that differ
in their mathematical meaning and chemical structure representation complement each other; in particular, prediction errors are mutually compensated.
Strategies for the Computer Prediction of the Properties of a Chemical Compound.
A strategy is an integral decision-making rule regarding the final activity of a chemical compound that is based on a set of intermediate prediction estimates of its
activity [96]. Methods of decision-making theory [54], particularly various voting
procedures [136], are used to develop such an integral decision rule. When classifying into two classes, a single activity estimate generated using a separate method
to describing a certain parameter group at a given level is a binary variable. Its
meanings correspond to one of alternate possibilities: “pro” or “contra.” Because
there are different classification methods, description levels and parameter groups,
the use of all of the intermediate calculated estimates of activity in the final vote
mimics an objective decision made by an independent expert group. The outcome
of this voting is context-independent with respect to both the method of structure
representation and the methods of intermediate prediction estimates construction,
and we can therefore regard these strategies as reliable tools for the evaluation of
untested compound activity.
Three prediction strategies are defined in IT Microcosm: a conservative strategy
that is based on a model of general unweighted consensus, a normal strategy that
uses a model of selective weighted consensus, and a risk strategy that implements a
model of supremum consensus.
The combined use of several different strategies during prediction constitutes a
universal hierarchic multistage final voting procedure; it mimics an objective decision by several independent expert groups. The ultimate integral decision rule is, in
essence, a multilevel consensus QSAR model.
P. M. Vassiliev et al.
2. No reduction of the dimensionality of mega-dimensional space of description
is carried out; the totality of the parameters is assumed to be informative. As a
consequence, no information determining the specifics of the objects to be recognized is lost, which allows for a more effective extrapolation of the obtained
regularities to new or poorly studied compounds with nontrivial specifics of
action.
3. A higher adequacy and enhanced prognostic capacity of the decision rules is
achieved by expanding the feature space by adding new parameter groups and
new levels of compound structure representation, rather than selecting “significant” variables or reducing the feature space.
4. This approach develops an ensemble of decision rules that are based on several
classification methods that differ essentially in their mathematical formalism
instead of a “better” prediction regularity. An independent classification of the
properties of the predicted object is done by each method for each description
level of each parameter group. The obtained spectrum of prediction estimates of
chemical compound activity is then used in final classification procedures.
5. The context-independence of the integral decision rule is a result of the generalization of the spectrum of intermediate prediction estimates of activity by the
methods of decision-making theory. As a result, primary decision rules that differ
in their mathematical meaning and chemical structure representation complement each other; in particular, prediction errors are mutually compensated.
Strategies for the Computer Prediction of the Properties of a Chemical Compound.
A strategy is an integral decision-making rule regarding the final activity of a chemical compound that is based on a set of intermediate prediction estimates of its
activity [96]. Methods of decision-making theory [54], particularly various voting
procedures [136], are used to develop such an integral decision rule. When classifying into two classes, a single activity estimate generated using a separate method
to describing a certain parameter group at a given level is a binary variable. Its
meanings correspond to one of alternate possibilities: “pro” or “contra.” Because
there are different classification methods, description levels and parameter groups,
the use of all of the intermediate calculated estimates of activity in the final vote
mimics an objective decision made by an independent expert group. The outcome
of this voting is context-independent with respect to both the method of structure
representation and the methods of intermediate prediction estimates construction,
and we can therefore regard these strategies as reliable tools for the evaluation of
untested compound activity.
Three prediction strategies are defined in IT Microcosm: a conservative strategy
that is based on a model of general unweighted consensus, a normal strategy that
uses a model of selective weighted consensus, and a risk strategy that implements a
model of supremum consensus.
The combined use of several different strategies during prediction constitutes a
universal hierarchic multistage final voting procedure; it mimics an objective decision by several independent expert groups. The ultimate integral decision rule is, in
essence, a multilevel consensus QSAR model.