Collaborative QSAR analysis of Ames mutagenicity.

We report the results of a collaborative QSAR modeling project between 15 teams to develop predictive computational QSAR models of in vitro Ames mutagenicity induced by organic compounds. The Ames dataset consisted of 6542 compounds (after curation). In total, 32 predictive classification QSAR models were developed using different combinations of chemical descriptors and machine learning approaches, representing the most extensive combinatorial QSAR modeling study ever done in the cheminformatics field in public domain. The resulting consensus model had the highest external predictive power nearly reaching the experimental reproducibility of 85% for the Ames test. In addition, we found published evidence indicating that 31 of 130 outliers (29 mutagens and 2 non-mutagens) were erroneously annotated in the original dataset. This work presents a model of collaboration that integrates the expertise of participating laboratories to establish the best practices and most reliable solutions for difficult problems in chemical and computational toxicology.