Study of structure–toxicity relationship by a counterpropagation neural network Original Research Article

The investigation presented here is an attempt to establish a model for the prediction of toxicity of molecules using artificial neural networks (ANN) with a counterpropagation learning strategy. Molecules have been described as 3D geometrical structures, i.e. by the (x, y, z)-coordinates of all atoms. Each structure has been encoded into a `spectrum-likeʹ representation as a suitable input for ANN modeling. As an extension to the coordinate related `spectrum-likeʹ representation, charge distributions, calculated with Mulliken population analysis, have been included in the modeling. A set of 41 benzene analogs were considered in this study, for which LD50 values were obtained from the literature. Several modeling experiments were performed on two training sets. All the models show good recall ability. The correlation coefficients of the models for retrieved vs. experimental values are larger than 0.9. The prediction ability of models is reasonable with correlation coefficients between 0.4 and 0.8. However, the quality of models depends on molecular representation and the choice of the training set.