Estimation of 13C-NMR chemical shift values using neural network technology

Vladimir Purtuc, Veronika Schütz,
Susanne Felsinger und
Wolfgang Robien



 
 
 
 
 

Neural networks represent an attractive tool for the prediction of physicochemical properties. We have focused our interest in the development of a general network allowing the prediction of C13-NMR chemical shift values for all classes of organic compounds.

The development of a neural network consists of several steps:

The data used during training and evaluation of the network are selected from the CSEARCH-NMR database holding some 230,000 carbon NMR spectra with a total number of 2,700,000 assigned chemical shift values. The severe restriction during the training, even when using the Alpha-Cluster, is based on the fact that the networks are comparably large depending on the number of molecular descriptiors selected, leading to a large number of weights to be optimized. Therefore a typical training set consists of only 400,000 examples selected on a random basis. The optimization of such a large network is an extremely time- and memory-consuming task, but the resulting parameter file has only a size of roughly 0.5MB holding the condensed information extracted from 230,000 carbon NMR spectra allowing a very fast and precise prediction of C-13 chemical shift values even at the PC-level.

The large advantage of our network design is the utilization of stereochemical information which further improves the quality of the prediction. The evaluation of stereochemical interactions is based on a technology with no need for 3-dimensional coordinates.

For a detailed description how to utilize this neural network for spectrum prediction using WEB-technology click here