ARTIFICIAL NEURAL NETWORK APPROACH FOR PREDICTION OF TOXICITY OF ORGANIC COMPOUNDS BASED ON AN IMPROVED GROUP CONTRIBUTION METHOD

摘要

A feed-forward artificial neural network (ANN) with an improved group contribution method has been developed to predict the aquatic toxicity (/GC50) of organic compounds to Tetrahymena pyriformis. In addition to the chemical groups, several molecular descriptors were applied to the ANN model to improve the predictive capability. The parameters of ANN models were divided into two parts (groups and properties) as the descriptors. A data set of 660 out of 824 organic compounds was used to train the ANN, while another data set of 82 organic compounds was used as a test set to validate the ANN model. The remaining 82 chemicals were used as the predicting set to prove that the ANN model can effectively predict the value of the toxicity. The network based on chemical groups had an accuracy of R~2=0.912 and MSE=0.090 while the improved group contribution method had an accuracy of R~2=0.948 and MSE=0.059.