C-13 NMR CHEMICAL-SHIFT SUM PREDICTION FOR ALKANES USING NEURAL NETWORKS

The C-13 NMR chemical shift sum (CSS) of alkanes was estimated with mu lti-linear regression (MLR) and multi-layer feed-forward artificial ne ural networks (ANN), using as structural descriptors the number of pat hs of length 1, 2, 3, and 4. The CSS prediction ability of both the ML R and ANN models was tested by the ''leave-20%-out'' (L20%O) cross-val idation method. Four activation functions were tested in the neural mo del: the hyperbolic tangent, a bell-shaped function, a linear function and the symmetric logarithmoid function. The linear and symmetric log arithmoid functions were used only for the output layer. All combinati ons of activation functions give close results both in calibration and cross-validation, with somewhat lower performances for the networks w ith a bell-shaped output function. The best results were exhibited by the networks with the symmetric logarithmoid output function, followed by the networks with a linear output function. Because the results we re very close, from a statistical point of view one could not definiti vely choose a particular combination of activation functions. The neur al model provides better calibration and cross-validation results than the MLR model. (C) 1998 Elsevier Science Ltd.