Toward an Optimal Approach for Variable Selection in Counter-Propagation Neural Networks: Modeling Protein-Tyrosine Kinase Inhibitory of Flavanoids Using Substituent Electronic Descriptors

Counter propagation neural network (CPNN) is one of the attractive tools of classification in QSAR studies. A major obstacle in classification by CPNN is finding the best subset of variables. In this study, the performance of some different feature selection algorithms including F score-based ranking, eigenvalue ranking of PCs obtained from data set, Non-Error-Rate (NER) ranking of both descriptors and PCs, and 3-way handling of data, Parallel Factor Analysis (PARAFAC), was evaluated in order to find the best classification model. The methods were applied for modeling protein-tyrosine kinase inhibitory of some flavonoid derivatives using substituent electronic descriptors (SED) as novel source of electronic descriptors. The results showed that the best performance was achieved by F-score ranking while the NER ranking of principal components (PCs) showed very fluctuate results and the worst performance was belonging to PARAFAC-CPNN. Furthermore, comparison of results of these nonlinear algorithms with linear discriminate analysis method revealed better predictions by the former.