Applicability Domain Characterization for Machine Learning QSAR Models

Abstract

AbstractQuantitative structure–activity relationship (QSAR) models serve as important tools for chemical hazard assessment. Recent decades witnessed an unprecedented success in developing “high-performance” QSAR models with various machine learning algorithms. Nonetheless, QSAR models are intrinsically data-driven models, in which patterns or rules are learned from training samples and thus can only be valid within limited applicability domains (AD). In order to be accepted for regulatory purposes, QSAR models should always be associated with defined ADs. In this chapter, essential concepts and understanding of AD first are introduced. Then, varied AD characterization methods and AD metrics are reviewed. Next, alterations to QSAR modeling scenarios resulted from the utilization of machine learning algorithms and potential influence of these alterations on the AD characterization methods are discussed. Finally, perspectives on AD characterization methods and development of QSAR models with broad AD are provided.