Artificial neural networks as versatile tools for prediction of MDR-modulatory activity

Following our ongoing studies on structure-activity relationship studies of propafenone-type modulators of multidrug resistance, we performed both a Free-Wilson analysis and a combined Hansch/Free-Wilson analysis on a set of 48 compounds using artificial neural networks (ANN). In comparison to classical multiple linear regression (MLR) analysis, the ANN showed equal or even slightly better predictive power in leave one out cross validation procedures and was remarkably superior when performing a leave 8 out cross validation. Additionally, it was possible to train a network using only 14 compounds and to properly predict the MDR-modulating activity of the remaining 34 compounds. In this case, the MLR analysis completely failed due to insufficient number of cases. Attempts to extract informations on which input descriptors are important using a genetic input selection algorithm failed. Best results were obtained using those descriptors which showed highest statistical significance in MLR analyses.