SIMULATION OF POLYSACCHARIDE C-13 NUCLEAR-MAGNETIC-RESONANCE SPECTRA USING REGRESSION-ANALYSIS AND NEURAL NETWORKS

A set of 13 polysaccharides was used as a training set to generate regression equations and to train back-propagation and quasi-Newton networks. These 13 polymers contained glucose, mannose, and xylose monomers with alpha and beta anomeric configurations linked at the (1 --> 3), (1 --> 4), and (1 --> 6) positions. An additional polysaccharide was utilized as a cross-validation set compound to test the external predictive ability of the regression equations and trained neural networks. The 10 descriptors used to represent the polysaccharides were identified through regression analysis. Although all three techniques yielded very accurate simulated spectra, quasi-Newton neural networks were found to yield superior results. The rms spectral error for the cross-validation set compound was 3.44, 1.43, and 0.71 ppm using regression analysis, back-propagation neural networks, and quasi-Newton neural networks, respectively. In addition, the number of cycles (epochs) required to train the quasi-Newton neural networks was significantly less than the number required to train back-propagation neural networks.