Application of Neural Network to Prediction of Structures of Labdane Diterpenes from 13C NMR Data

Abstract

This study applies Generalized Regression Neural Network to elucidate the structures of novel Labdane diterpenes. Firstly, skeletal 13 C values for the labdane skeleton were predicted. Secondly, the substituents attached to each of the positions were predicted. In predicting the skeletal values for the unknown compounds, the 13 C data of 121 Labdane diterpenes were used as input to GRNN while the corresponding skeletal 13 C values were used as the target data. After training, the network was simulated using 25 test compounds. This network was referred to as GR_SC. Two networks (GR_LAD and GRSC_LAD) were used to predict the substituents on the labdane skeleton. In GR_LAD, the skeletal data of the 121 compounds were used as training data while the coded substituents were used as the target data. In GRSC_LAD, the skeletal 13 C data for 121 compounds were used as the input data. The chemical shift ranges over which each substituent type may be obtained were used as the target data. In both procedures, the network was subsequently trained and simulated using the 13 C data predicted by GR_SC. Best results were obtained at spread constants of 3.0, 3.0 and 0.1 respectively for GR_SC, GR_LAD and GRSC_LAD respectively. The chemical shift values predicted by GR_SC were close to the actual values. GR_LAD predicted only the most likely substituent for each position on the skeletons of the unknown labdane diterpenes while GRSC_LAD gave all the possible substituents for each position. Both methods may be helpful in predicting the structures of unidentified substances.