Using neural networks for (13)c NMR chemical shift prediction-comparison with traditional methods

Interpretation of (13)C chemical shifts is essential for structure elucidation of organic molecules by NMR. In this article, we present an improved neural network approach and compare its performance to that of commonly used approaches. Specifically, our recently proposed neural network (J. Chem. Inf. Comput. Sci. 2000, 40, 1169-1176) is improved by introducing an extended hybrid numerical description of the carbon atom environment, resulting in a standard deviation (std. dev.) of 2.4 ppm for an independent test data set of approximately 42,500 carbons. Thus, this neural network allows fast and accurate (13)C NMR chemical shift prediction without the necessity of access to molecule or fragment databases. For an unbiased test dataset containing 100 organic structures the accuracy of the improved neural network was compared to that of a prediction method based on the HOSE code (hierarchically ordered spherical description of environment) using SPECINFO. The results show the neural network predictions to be of quality (std. dev. = 2.7 ppm) comparable to that of the HOSE code prediction (std. dev. = 2.6 ppm). Further we compare the neural network predictions to those of a wide variety of other (13)C chemical shift prediction tools including incremental methods (CHEMDRAW, SPECTOOL), quantum chemical calculation (GAUSSIAN, COSMOS), and HOSE code fragment-based prediction (SPECINFO, ACD/CNMR, PREDICTIT NMR) for the 47 (13)C-NMR shifts of Taxol, a natural product including many structural features of organic substances. The smallest standard deviations were achieved here with the neural network (1.3 ppm) and SPECINFO (1.0 ppm).