In the doctoral thesis, we focused on the investigation of computer models for the prediction of endocrine disruptors by binding to selected nuclear receptors (androgen, estrogen, glucocorticoid and thyroid hormone receptors). First, we investigated whether freely available computer models can adequately predict the binding ability of a selected group of compounds. The predictions for the selected group of compounds were compared with in vitro data collected in the CompTox Chemistry Dashboard database. We found that although the existing selected computer models had similar prediction accuracy, they were not accurate enough to be used as stand-alone tools. Therefore, we tried to achieve more useful results by combining the models. Although the three combined models did not improve accuracy, they provided more useful results. Prediction with the innovative negative consensus rule model significantly improved sensitivity. A significant reduction in false negatives allows efficient identification of inactive compounds that can be excluded from further testing at other levels. In the second part of the research, we developed our own computer models using artificial neural networks to predict the binding of compounds to selected nuclear receptors as agonist and antagonist separately. We evaluated the accuracy of the models by cross-validation. The predictive power of the developed models is high as the accuracy of binding prediction is significant. Furthermore, by adjusting the classification threshold of the models, we can effectively identify the compounds for which the prediction uncertainty is the highest. These compounds are included in further testing, and a refined prediction is made for the remaining compounds. We have also examined structural fragments of compounds that bind to several selected nuclear receptors. Compounds that contain a nitrogen atom attached to a chain of at least eight carbon atoms and that do not contain oxygen are likely to bind to the selected nuclear receptors.
|
