379
12 Consensus Drug Design Using IT Microcosm
A Complex Methodology for the Computer Prediction of the Properties of a
Chemical Compound. This is the central concept of IT Microcosm [93, 105, 116]
and features a synthesis of the fundamental theoretical concepts that underlie the
development of rules and principles for generating the applied components of the
technology: algorithms and programs that predict compound activity.
An adequate prediction of chemical compound activity is only possible through
the generalization of a spectrum of prediction estimates that are obtained by several
methods of classification that differ in their mathematical formalism; these methods
are applied to levels of varying complexity and methods of structure representation
with a varying physicochemical meaning. We use every available description variable and the redundancy of this description when expanding parameters, on the basis of several decision-making circuits that are conceptual in the results they yield.
The following decision making circuit conforms to the concept described above:
1. constructing a representative training set that includes the structures of reliably
active and inactive tested substances;
2. constructing models of the generalized patterns of active/inactive compound
classes on the basis of a mega-dimensional multilevel description of their structure with parameter groups of different physicochemical meaning;
3. establishing a set of decision rules using several essentially different classification methods; each method is used separately for each description level of each
parameter group;
4. calculating the spectrum of prediction estimates of compound activity in the
training set using all of the developed classification rules; and
5. constructing integral multimodel consensus decision-making rules using strategies with different spectra of prediction estimates, and evaluating the prognostic
ability of these decision rules.
Utilizing this complex methodology, we obtain decision rules and results of activity
prediction that are context-independent from the training set composition, the methods of compound structure representation, and the methods of regularity detection.
The following principles of constructing the applied components of IT Microcosm arise from the summation of these theoretical concepts:
1. a compound structure should be described by the maximum possible number of
parameters;
2. the structure representation should be multilevel;
3. the groups of description parameters should differ in their physicochemical
meaning;
4. the decision rules for activity prediction should include all of the parameters of
structure representation;
5. the prediction methods should be adapted for use in nonlinear spaces with extralarge dimensions;
6. the compound activity prediction should be performed by several methods
simultaneously;
7. the prediction methods should differ in their mathematical formalism; and
12 Consensus Drug Design Using IT Microcosm
A Complex Methodology for the Computer Prediction of the Properties of a
Chemical Compound. This is the central concept of IT Microcosm [93, 105, 116]
and features a synthesis of the fundamental theoretical concepts that underlie the
development of rules and principles for generating the applied components of the
technology: algorithms and programs that predict compound activity.
An adequate prediction of chemical compound activity is only possible through
the generalization of a spectrum of prediction estimates that are obtained by several
methods of classification that differ in their mathematical formalism; these methods
are applied to levels of varying complexity and methods of structure representation
with a varying physicochemical meaning. We use every available description variable and the redundancy of this description when expanding parameters, on the basis of several decision-making circuits that are conceptual in the results they yield.
The following decision making circuit conforms to the concept described above:
1. constructing a representative training set that includes the structures of reliably
active and inactive tested substances;
2. constructing models of the generalized patterns of active/inactive compound
classes on the basis of a mega-dimensional multilevel description of their structure with parameter groups of different physicochemical meaning;
3. establishing a set of decision rules using several essentially different classification methods; each method is used separately for each description level of each
parameter group;
4. calculating the spectrum of prediction estimates of compound activity in the
training set using all of the developed classification rules; and
5. constructing integral multimodel consensus decision-making rules using strategies with different spectra of prediction estimates, and evaluating the prognostic
ability of these decision rules.
Utilizing this complex methodology, we obtain decision rules and results of activity
prediction that are context-independent from the training set composition, the methods of compound structure representation, and the methods of regularity detection.
The following principles of constructing the applied components of IT Microcosm arise from the summation of these theoretical concepts:
1. a compound structure should be described by the maximum possible number of
parameters;
2. the structure representation should be multilevel;
3. the groups of description parameters should differ in their physicochemical
meaning;
4. the decision rules for activity prediction should include all of the parameters of
structure representation;
5. the prediction methods should be adapted for use in nonlinear spaces with extralarge dimensions;
6. the compound activity prediction should be performed by several methods
simultaneously;
7. the prediction methods should differ in their mathematical formalism; and