The prediction of carbon-13 NMR chemical shifts using ensembles of networks

Ensembles of a multilayer network are set up to predict the carbon-13 nuclear magnetic resonance (C¹³ NMR) chemical shifts of a series of monosubstituted benzenes. The descriptors (inputs) used are twelve structural-based vectors that correspond to the calculated Huckel and Gasteiger electron densities of the monosubstituted aromatic systems and four graphical descriptors that correspond to the numbers, of appearance of some specific structural features of the substitutents. The outputs are the C¹³ NMR chemical shifts of the ipso, ortho, meta, and para carbons. A training set of 38 data was used and, after training, the neural network was tested for its ability to predict the C¹³ NMR chemical shirts of 15 compounds not included in the training set. The authors demonstrated that the performance of artificial neural networks in C¹³ NMR chemical shift prediction could be improved by (a) using both structural-based and graphical descriptors as input parameters, (b) pruning, and (c) combining the prediction from a number of networks. Furthermore, pruning the connection weights can also enable one to select the appropriate input variables