Feature-selection techniques and modeling approaches in QSAR and QSPR.

A Quantitative Structure-Activity Relationship (QSAR) is a linear or nonlinear model, which relates variations in molecular descriptors (containing information about the structure) to variations in biological activity using a series of active and/or inactive compounds. It can either be used for prediction of the activity of untested drugs or to set up a priority plan of synthesis and experimental testing of new compounds. There are several steps which are important to build a QSAR/QSPR models. For instance, the feature selection and the modeling are two of them. The selection of the relevant molecular descriptors from more than one thousand calculated descriptors is a crucial step in the development of QSAR models. In spite of the fact that many feature-selection methods already have been applied, it is still unclear which to use best to find a stable and reliable feature selection for a given type of data set. The main aim of the thesis was to compare different feature-selection and modeling techniques in QSAR/QSPR studies.