Automatic interpretation of nucleoside mass spectra

Various heuristic and pattern recognition techniques have been applied to a data base of 125 underivatised nucleoside mass spectra to determine certain aspects of structure from an unknown spectrum. A heuristic program has been written encoding nucleoside mass spectral fragmentations in order to determine molecular weight, formula weight of the purine or pyrimidine base part, and, unsuccessfully, base type. The pattern recognition methods of statistical linear discriminant function analysis, learning machine approach, distance from the mean, and k-nearest neighbour classification have been applied to the same data base divided into training and prediction sets. Analyses were conducted to determine numbers of carbon, oxygen, and nitrogen atoms present in the base alone and in the nucleoside as a whole, substitution patterns, and base type. Prediction success for all approaches was typically in the range 76-86%, or in terms of the figure of merit 0.20 - 0.27