Pattern recognition and interpretation of nucleoside mass spectra

Abstract

Four pattern-recognition techniques-distance from the mean, the learning machine approach, a statistical linear discriminant function analysis and the k-nearest-neighbour method-have been applied to a set of 125 nucleoside mass spectra. Twenty-one structural features, comprising elemental compositions and substitution types, were predicted from binary and logarithmic transforms of the spectra. The best classification was given by the distance from the mean method with logarithmic data. With this method prediction success averaged 79 % over the 21 categories, or in terms of the figure of merit 0.27.