Prediction of antiprion activity of therapeutic agents with structure-activity models.

We have developed computational structure-activity models for the prediction of antiprion activity of compounds with known molecular structure. The aim is to apply the developed classification and predictive models in further drug design of antiprion therapeutics. The neural network models developed on the counter-propagation reinforcement learning strategy performed better than the linear regression models. The initial data set was composed of 461 compounds representing diverse groups of chemicals (derivatives of acridine, quinolone, Congo red, 2-aminopyridine-3,5-dicarbonitrile, styrylbenzoazole, 2,5-diamino-benzoquinone), which have been tested in comparable cell-screening assay studies for their activity against prion accumulation. Initially, we have designed a classification model for preliminary sorting of compounds into highly active, active, and inactive groups. Further, only the active compounds with IC50 less or equal to 10 μM were considered as the initial source of data. Altogether, 158 compounds were used to train the artificial neural network model for the estimation of the antiprion activity. The predictive ability of the model was significantly improved after selection of influential variables with genetic algorithm. The root- mean-squared error of the predicted pIC50 values for the external validation set (RMS EV) was slightly above 0.50 log units. A linear regression model, developed for the reasons of comparison, performed with a lower predictive ability (RMS EV 0.92 log units). The applicability domain of the models was assessed by a leverage and distance approach. The set of selected influential structural variables was further studied with the aim to get a better insight into the structural features of compounds potentially involved in disturbing of the prion-prion interactions.