Retrospective (Q)SAR analysis of genotoxicity data from European pharmaceutical scientific advice and authorization procedures and performance improvements using structural alert and chemical subspace analyses, reliability filtering, and expert review.

Drug development warrants a high number of animal studies, fostering the development of New Approach Methodologies (NAMs) to reduce, replace and refine animal studies (3Rs). As part of this effort, various (Q)SAR models have been developed, and many of these models were designed to predict genotoxicity. To retrospectively investigate the performance and regulatory applicability of available computational genotoxicity prediction models and to evaluate methods for improving prediction performances, a dataset was manually curated from proprietary and publicly available Austrian, French, and European regulatory pharmaceutical procedures. The collected experimental endpoints (Ames bacterial mutagenicity, in vitro mammalian cell genotoxicity, and in vivo genotoxicity) were then retrospectively analyzed with different available (Q)SAR prediction models. Model performance proved to be endpoint and model specific, with Ames test models routinely used in regulatory ICH M7 analyses showing better performances than in vitro mammalian cell and in vivo genotoxicity models. In many cases improvements of prediction performance were possible by applying structural alert analyses, chemical subspace analyses, prediction reliability filtering, and expert review. Still, in vitro mammalian cell genotoxicity and in vivo genotoxicity models need further refinement before they can be applied in regulatory decision making.