Recent reports that a wide variety of natural and man-made compounds are capable of competing with natural hormones for estrogen receptors serve as timely examples of the need to advance screening techniques to support human health and ascertain ecological risk. Quantitative structure-activity relationships (QSARs) can potentially serve as screening tools to identify and prioritize untested compounds for further empirical evaluations. Computer-based QSAR molecular models have been used to describe ligand-receptor interactions and to predict chemical structures that possess desired pharmacological characteristics. These have recently included combined and differential relative binding affinities of potential estrogenic compounds at estrogen receptors (ER) alpha and beta. In the present study, artificial neural network (ANN) QSAR models were developed that were able to predict differential relative binding affinities of a series of structurally diverse compounds with estrogenic activity. The models were constructed with a dataset of 93 compounds and tested with an additional dataset of 30 independent compounds. High training correlations (r2=0.83-0.91) were observed while validation results for the external compounds were encouraging (r2=0.62-0.86). The models were used to identify structural features of phytoestrogens that are responsible for selective ligand binding to ERalpha and ERbeta. Numerous structural characteristics are required for complexation with receptors. In particular, size, shape and polarity of ligands, heterocyclic rings, lipophilicity, hydrogen bonding, presence of quaternary carbon atom, presence, position, length and configuration of a bulky side chain, were identified as the most significant structural features responsible for selective binding to ERalpha and ERbeta.