The receptor-like neural network for modeling corticosteroid and testosterone binding globulins.

Abstract

A neural-net method for simulation of corticosteroid and testosterone binding globulin (CBG, TBG)-ligand interactions is presented. Molecular modeling provides the geometry and partial atomic charges of 31 steroid molecules. The atomic coordinates within the molecule of the compound of the highest affinity are then used to train a self-organizing map (SOM) that forms a template for the comparison to other molecules. Comparison is done using a series of normalized patterns produced by the SOM. The template SOM, after overlaying on the set of random vectors, mimics the topology of the receptor site and is used to train unsupervisedly a neuron capable of recognizing the degree of similarity between the reference and tested patterns. A good correlation is observed for signals generated by the neuron plotted against the experimental CBG affinities. For TBG affinity modeling a modified procedure is designed which is capable of separating electrostatic and shape effects. The high predictive power of the model is achieved by keeping close analogy to the processes taking place at the real receptor sites.