Optimization of automatically generated rules for predicting the presence and absence of substructures from MS and MS/MS data

Abstract

A pattern-recognition/artificial-intelligence program, referred to as MAPS (Method for Analyzing Patterns in Spectra), was recently developed to identify the relationships that exist between substructures and the characteristic features they produce in the spectra from mass spectrometry (MS) and successive mass spectrometry (MS/MS). MAPS has been extended to utilize these relationships to formulate exclusion rules as well as inclusion rules, so that the absence of recognized substructures can be predicted as well as their presence. The potential usefulness of each MS and MS/MS spectral feature in such rule formulation is characterized by correlation and uniqueness factors. The correlation factor expresses the degree of correlation between a feature and a specific substructure; the uniqueness factor expresses the uniqueness of a feature with respect to that substructure. Features with high correlation factors are most use for predicting the absence of substructures, whereas features with high uniqueness factors are most useful for predicting their presence. Feature intensity-data have been found to improve the inclusion-rule performance and degrade the exclusion-rule performance. Criteria for optimizing the predictive abilities of both rule types are discussed.