MADALINA HRUBARU, LASZLO TARKO A NOVEL METHOD OF VALIDATION OF THE QSPRS BASED ON MOLECULAR SIMILARITY For practical drug design purposes, the proposed method should be applied only if the analyzed database includes a prediction set (a group of a new molecules, not yet synthesized, with unknown values of the dependent property). The characteristics of the proposed method are: a) elimination of some molecules from the initial calibration and prediction sets, according to the result of a specific molecular similarity procedure b) the validation set is identified based on the results of similarity calculations and includes the molecules of the new calibration set most similar to the molecules of the new prediction set c) inclusion of the validation set in the new calibration set, used for model building d) it uses an original mathematical formula as validation function e) the most suitable equation for the description of the molecules in the new calibration set is different from the validated equation; these two equations are identified within the group of the best 1000 QSPRs f) the validated equation is considered the most suitable equation for the description of the molecules in the new prediction set. In five QSPR studies the validated equation made better prediction for molecules in the new prediction set than the most suitable equation for the description of molecules in the new calibration set. In the sixth study the proposed method produced, for the same prediction set, a better result than an external validation method.
Keywords: drug design, molecular similarity, QSPR / QSAR, validation