IVL Swedish Environmental Research Institute

Hormone regulated processes govern the larval development and reproduction in aquatic vertebrates. The involvement of pollutants in these processes needs to be examined in the evaluation of consequences of their release into the aquatic environment. Prognosis models may be used as a pre-screening step to predict such properties. This work describes the development of models to predict the properties involved in estrogenic hormone activity, i.e. induction (in absolute and relative figures) and binding affinity to the human estrogen receptor Models were developed using molecular hologram descriptors, i.e. derived only from the chemical structure of substances. Statistical molecular design (SMD) was used to select substances for experimental testing of receptor induction. Substances were then divided in a set for model construction and a validation set. Models were calculated for a large number of systematically varied hologram configurations to find those with the best data based on the predictive ability (Q2-value). The models with the best predictive ability were those based on holograms that considered chirality in the chemical structure. No other systematic effect of hologram length and fragment length was observed. The best models for each biological response were then further refined by a pruning procedure that resulted in exclusion of descriptors that did not contribute positively to the model. The three estrogenic responses were predicted within a factor 10 (root mean square error of prediction 0.7 - 1.0), which should be sufficient for the pre-screening purpose. A broad applicability characterises the three models, only the structural formula of a new substance is needed to be able to perform the prediction.

Hormone regulated processes govern the larval development and reproduction in aquatic vertebrates. The involvement of pollutants in these processes needs to be examined in the evaluation of consequences of their release into the aquatic environment. Prognosis models may be used as a pre-screening step to predict such properties. This work describes the development of models to predict the properties involved in estrogenic hormone activity, i.e. induction (in absolute and relative figures) and binding affinity to the human estrogen receptor Models were developed using molecular hologram descriptors, i.e. derived only from the chemical structure of substances. Statistical molecular design (SMD) was used to select substances for experimental testing of receptor induction. Substances were then divided in a set for model construction and a validation set. Models were calculated for a large number of systematically varied hologram configurations to find those with the best data based on the predictive ability (Q2-value). The models with the best predictive ability were those based on holograms that considered chirality in the chemical structure. No other systematic effect of hologram length and fragment length was observed. The best models for each biological response were then further refined by a pruning procedure that resulted in exclusion of descriptors that did not contribute positively to the model. The three estrogenic responses were predicted within a factor 10 (root mean square error of prediction 0.7 - 1.0), which should be sufficient for the pre-screening purpose. A broad applicability characterises the three models, only the structural formula of a new substance is needed to be able to perform the prediction.