An algorithm for the selection of best multivariate regression models (according to correlation coefficient) was developed for the use in drug research. It is realized in the programming language Visual Basic and connected to a database. This work explores the relationship between the statistical parameters calculated in the process of fitting on the training data set, cross-validation by dropping a certain percentage of data study sets (k-fold cross-validation), and in the process of prediction on an external data set. Cross-validation procedure, in which a larger sub-set of data is omitted in each step, shows better agreement with results obtained in prediction on external set than the leave-one-out cross-validation procedure. Additionally, correlations between statistical parameters calculated in the fitting procedure on the training set and corresponding parameters calculated in the cross-validation procedures and in validations on external data sets are much better than it was published in literature. Such an analysis confirmed very good attributes of applied procedures for initial elimination of descriptors, reduction of inter-correlations between descriptors, and very good attributes of modeling method and algorithm used for model selection. Unlike a regular cross-validation procedures used in the literature, frequencies of occurences of individual descriptors in the best models were explored. Based on this analysis an order of importance of individual descriptors in best models was made. This information is important in the interpretation and use of best models. Obtained results and developed application MR_QSAR are valuable scientific and technical contribution to the development and application of computer algorithms for modeling molecular properties and activities carried out in scientific circles, in studies of new drugs in the pharmaceutical industry, and in environmental protection in procedures for assessment the toxicity of the molecules.