ABSTRACT
A methodology is presented for modeling and classifying endocrine disruptors based on Kohonen and counterpropagation neural networks. Three different datasets were considered. The rst dataset consists of 106 substances extracted from the list of 553 chemicals that were inspected by the European Union Commission for the scientic evidence of their endocrine disruption activity. For this dataset, we present the classication model designed for a preliminary assessment of potential endocrine disruptors, which would help the assessors to make the priority list for a large amount of chemicals that have to be tested with more expensive in vitro and in vivo methods. The second dataset consists of 132 compounds of known chemical structures, which were tested for their binding afnities to the mice estrogen receptor. We compared the counterpropagation neural network models for the prediction of relative binding afnity with two other multivariate modeling methods (partial least square regression, and error-back-propagation neural network. The results were assessed with the aim to get insight into the mechanisms involved in the binding of estrogenic compounds to the receptor. The third dataset encompasses 60 diverse chemicals tested for the binding afnity to human estrogen receptors a and b (ER-a and ER-b). To obtain the structure-activity relationship, the three-dimensional (3D) structures of ligands and receptors were taken into consideration. Structural features of ligands having the strongest inuence to the binding afnities were investigated.
