Chapter 7 Variable Selection QSAR Modeling, Model Validation, and Virtual Screening

Abstract

Publisher Summary This chapter discusses and illustrates variable selection that rigorously validate quantitative structure–activity relationship (QSAR) models with statistically significant external predictive power ultimately afford their application as reliable virtual screening tools for database mining or chemical library design. Several criteria of model robustness have been presented in the chapter, particularly emphasizing on the external predictive power of the model. The traditional LOO q 2 is insufficient statistical characteristic of model predictive power. Thus, although low values of q 2 almost certainly indicate that the model is not expected to be predictive externally, high values of this parameter alone also provide no guarantee that the underlying models can be used for the reliable external prediction. The application of these principles to the development of the predictive QSAR models of anticonvulsant agents followed by the use of these models for database mining is shown to afford the discovery of novel anticonvulsant agents. The robustness and external predictive power and applicability domain as the most important parameters of QSAR models shall increase their practical use as reliable virtual screening tools for pharmaceutical lead identification.