To date the use of artificial neural networks (ANNs) in quantitative structure-activity relationship (QSAR) studies has been primarily concerned in comparing the predictive accuracy of the technique using known data sets where the data set parameters had been preselected and optimized for use with other statistical methods. Little effort has been directed at optimizing the input parameters for use with ANNs or exploring other potential strengths of ANNs. In this study, back-propagation ANNs and multilinear regression (MLR) were used to examine the QSAR between substituent constants and random noise at six positions on 57 1,4-benzodiazepin-2-ones (1,4-BZs) and their binding affinities (log IC50) for benzodiazepine GABAA receptor preparations. By using selective pruning and cross-validation techniques, it was found possible to use ANNs to indicate an optimum set of 10 input parameters from a choice of 48 which were then used to train back-propagation ANNs that best predicted the receptor binding affinity with a high correlation between known and predicted data sets. Using the optimum set of input parameters, three-layer ANNs performed no better than the two-layer ANNs which gave marginally better results than MLR. Using the trained ANNs to examine the individual parameters showed that increases in the lipophilicity and F polar value at position 7, F polar value at position 2', and dipole at position 1 on the molecule all enhanced receptor binding affinity of 1,4-BZ ligands. Increases in molar refractivity and resonance parameters at position 1, molar refractivity at positions 6' and 2', Hammet meta constant at position 3', and Hammet para constant at position 8 on the molecule all caused decreases in receptor binding affinity. By considering the optimal ANNs as pharmacophore models representing the internal physicochemical structure of the receptor site, it was found that they could be used to critically examine the properties of the receptor site.