Unsupervised Chemical Pattern Recognition In Complex Mass Spectra

Abstract

Developing a reliable method for obtaining chemical information on components of complex materials is a major goal of modern pyrolysis mass spectrometry. Unsupervised methods are necessary since pure reference spectra of components are often unavailable. Recent developments in unsupervised mixture analysis using examples from bacteria, jet fuels, and wood are given. Factor and discriminant analysis on 18 strains of Escherichia coli showed a clear differentiation between K1 and non-K1 antigen strains. Graphical rotation confirmed that the differentiation was based on the K1 antigen colominic acid. Binary and ternary mixtures of jet fuel components were used to numerically extract spectra of pure compound classes present in fuel formulations. Mathematically extracted spectra compared favorably to component spectra, which were not included in the data set. The variance diagram technique gave the same directions of the component axes as target rotation, a supervised technique. Variance diagram methods were applied to single samples of wood by recording mass spectra as a function of temperature, since the three major components of wood pyrolyze at slightly different temperatures. A variance diagram with factor 2 expanded showed clearly the three major wood components: cellulose, hemicellulose, and lignin.