PREDICTION OF C-13 NUCLEAR-MAGNETIC-RESONANCE CHEMICAL-SHIFTS BY ARTIFICIAL NEURAL NETWORKS

Empirical models relating atom-based structural descriptors to C-13 NMR chemical shifts have been used to accurately simulate C-13 NMR spectra for compounds whose shifts are unknown. In this work neural networks are investigated as a supplement to regression analysis in linking structural descriptors to chemical shifts. A recently studied data set of 431 keto-steroid carbon atoms is reexamined using neural networks. This approach allows the neural network aspects of the study to be emphasized and provides a basis for comparison with regression results. A fully-connected three-layer neural network was trained to predict C-13 NMR chemical shifts with 0.1 ppm resolution using the back-propagation learning algorithm. A cross-validation set was used to monitor the training and to prevent overfitting the data. The results using the neural network approach had standard errors about half as large as those achieved with linear regression techniques. In addition to providing more accurate predictions, the neural network approach also reports a confidence value for each observation. The results obtained with the neural networks are dependent on the initial conditions prior to training. For this reason, the neural network experiments were repeated a number of times with varying starting conditions, network architectures, and random cross-validation subsets. In every cass examined, the neural network results were superior to those achieved using linear regression analysis.