Towards accurate Virtual Screening Technologies: Sampling of Chemical Space and Applicability Domain Estimation.

A. TALEVI; E. A. CASTRO; L. E. BRUNO-BLANCH

Chemical information and computational challenges in XXI Century

Nova Science publishers

Año: 2012; p. 277 - 297

As QSAR modeling becomes more and more used in almost all the fields of chemistry and biology, assuring the reliability of QSAR models predictions has turned into a matter of most importance, especially if we take into account that, for the first time in history, QSAR models are being applied to replace experimental testing (in regulation contexts) in the field of chemical risk assessment. Adequate sampling of the chemical space (either to design a training set or to obtain a rational partition of a dataset into training and test samples), proper validation, and applicability domain estimation are the three main tools available for the modeler in order to assure predictions reliability. Here we overview recent advances and applications in this field. We analyze different sampling approaches designed to extract representative points from a given chemical space (factorial, space filling, D-optimal, onion designs), hints to avoid over-fitting emerging from recent studies on validation procedures, and a summary of different applicability domain assessment methodologies (distance-based, similarity-based, parametrical, non-parametrical, and others).