There are two broadly-defined applications of artificial neural networks (ANNs) in SAR/QSAR modeling. The first is the use of ANNs as prespocessors in order to transform descriptors into a form amenable to statistical analyses, such as regression modeling or linear discriminant analysis. This class of application often uses self-organizing feature map (SOFM) networks. For example, SOFMs are used to reduce three-dimensional spatial descriptors to two-dimensional matrices. The second major application of neual networks is as a statistical tool with which to classify chemicals by activity under a given biological end point. The neural network classification models are analogous to those derived from binary logistic regression (LR) or linear discriminant analysis (LDA). Like its statistical counterparts, a neural network can be developed for use in predictive toxicology. Our group has been using neural networks primarily as classifiers for problems in predicitive SAR/QSAR modeling. Case studies have been derived from data bases consisting of chemicals calssified as carcinogens, mutagens or under various non-cancer end points. Our analyses using neural networks are typically compared side-by-side with LR and LDA. Results obtained thus far show that ANNs can perform on an equal footing with the statistical tools when measured in terms of overall predictivity or the models. While the ANNs are easy to adapt to SAR/QSAR problems, the technique is limited by a lack of developed methodology for selecting the best variable subsets for the model: basic modeling diagnostics and network pruning procedures are often performed manually, as opposed to the "automatic" procedures in best subsets or stepwise regression techniques. Ultimately, combinations of neural networks and statistical methods might prove the be the best approach to predictive SAR/QSAR modeling. Examples of ANN aaplications will be presented and compared with models from statistical techniques.