Molecular surface properties such as the electrostatic or the hydrophobicity potential were condensed into an autocorrelation descriptor. A vector of these autocorrelation descriptors based on the molecular electrostatic potential was successfully applied to modeling the affinities of a set of 31 steroid molecules binding to the corticosteroid binding globulin (CBG) receptor by using a combination of a Kohonen and a feedforward neural network. Similarly, an autocorrelation vector derived from the hydrophobicity potential was used to model the binding constant of a set of 78 polyhalogenated aromatic compounds to the cytosolic Ah receptor. The models found have a high predictive ability as established by cross-validation.