engleski

CROMRSEL-s : efficient algorithms for the selection of most important variables in the QSAR modeling

CROMRsel is a suite of programs for variables selection based on the multiregression (MR) models. Three algorithms were developed: (1) for the selection of the best possible MR models containing I descriptors selected from the total number of N descriptors; (2) for the stepwise 'i by i' selection (usually, i =1-4) - starting from the small initial set of descriptors (selected by algorithm 1), in each next step i descriptors were added to those selected in the previous step; (3) variant of (2) in which N models containing I descriptors were generated in such a way that stepwise selection starts from each of N descriptors and, finally, the best among N models is chosen. For the cases in which N!/(N-I)!×I! < 1010 algorithm (1) and, otherwise, algorithms (2) and (3) was used. CROMRsel algorithms were applied to several data sets (studied by NNs, PLS, PCA, genetic algorithms or classical regression) from medicinal, analytical and physical chemistry, and, in all cases, we obtained significantly better models.

variable selection; multiregression; neural network; modeling

nije evidentirano