The 3D structures of all molecules studied in the investigations reported here were generated by the 3D structure generator CORINA [6]-[9] and used without further optimization. Partial atomic charges were calculated by the PEOE (Partial Equalization of Orbital Electronegativity) method [10] and its extension to conjugated systems [11]. Residual electronegativity values were also obtained by the above two procedures [10][11] by considering the charge dependence of electronegativities [12]. The method for the calculation of atom polarizabilities has been published in reference [13]. These methods are collected in the program package PETRA (Parameter Estimation for the Treatment of Reactivity Applications) [14].
The molecular electrostatic potential was obtained in a classical manner by moving a point charge on the molecular surface and calculating the potential according to Coulomb’s law from the partial atomic charges [15]. Any molecular surface can be taken into account; in most cases, we used the van der Waals surface.
The Kohonen neural networks were generated and analyzed by the Kohonen map simulator KMAP [16]. The study of the benzodiazepine and dopamine dataset was performed with an implementation of a Kohonen network on a massively parallel computer, MasPar [17][18].
In many studies reported here, the same dataset was used: 31 steroids binding to the corticosteroid binding globulin (CBG) receptor [19][20] and for which affinity data were available in the literature [21]-[23]. This dataset was chosen because it had been studied with other methods [20][24][25]. Quite intentionally the same dataset of CBG steroids is used again and again, in order to render the various methods comparable and show which features they emphasize.