The study and prediction of kinase function (kinomics) is of major importance for proteome research due to the widespread distribution of kinases. However, the prediction of protein function based on the similarity between a functionally annotated 3D template and a query structure may fail, for instance, if a similar protein structure cannot be identified. Alternatively, function can be assigned using 3D-structural empirical parameters. In previous studies, we introduced parameters based on electrostatic entropy (Proteins 2004, 56, 715) and molecular vibration entropy (Bioinformatics 2003, 19, 2079) but ignored other important factors such as van der Waals (vdw) interactions. In the work described here, we define 3D-vdw entropies (degrees theta(k)) and use them for the first time to derive a classifier for protein kinases. The model classifies correctly 88.0% of proteins in training and more than 85.0% of proteins in validation studies. Principal components analysis of heterogeneous proteins demonstrated that degrees theta(k) codify information that is different to that described by other bulk or folding parameters. In additional validation experiments, the model recognized 129 out of 142 kinases (90.8%) and 592 out of 677 non-kinases (87.4%) not used above. This study provides a basis for further consideration of degrees theta(k) as parameters for the empirical search for structure-function relationships.

Variable selection is a procedure used to select the most important features to obtain as much information as possible from a reduced amount of features. The selection stage is crucial. The subsequent design of a quantitative structure-activity relationship (QSAR) model (regression or discriminant) would lead to poor performance if little significant features are selected. In drug design modern era, by the means of combinatorial chemistry and high throughput screening, an unprecedented amount of experimental information has been generated. In addition, many molecular descriptors have been defined in the last two decays. All this information can be analyzed by QSAR techniques using adequate statistical procedures. These techniques and procedures should be fast, automated, and applicable to large data sets of structurally diverse compounds. For that reason, the identification of the best one seems to be a very difficult task in view of the large variable selection techniques existing nowadays. The intention of this review is to summarize some of the present knowledge concerning to variable selection methods applied to some well-known statistical techniques such as linear regression, PLS, kNN, Artificial Neural Networks, etc, with the aim to disseminate the advances of this important stage of the QSAR building model.