Machine Intelligence Applied to Chemical Systems: A Graph Theoretical and Learning Machine Study of Second-Order Effects in Low Resolution Mass Spectra

Abstract

Low resolution mass spectra can be usefully classified using pattern classifiers based on linear threshold logic units. These systems are linear in that they use the mass spectrometry peaks independently of one another. However, the theory of mass spectrometry as well as pattern classification considerations suggests that second-order interactions (cross terms which consider relationships between peaks) could be used to advantage in performing such classifications. A similarity measure has been used to develop two types of cross terms (intraset and interset cross terms) from low resolution mass spectra. It is shown that the interset cross terms thereby derived have a high probability of being correlated with the molecular features which define the categories of classification. The method has been implemented with threshold logic unit pattern classifiers working with several sets of mass spectrometry data. The cross terms are shown to increase the power of the pattern classification systems by speeding up convergence and/or raising their predictive ability.