Identification of Putative Skin Sensitizers Using QSAR Models
Alexander Tropsha, Vinícius M. Alves, Eugene N. Muratov, Denis Fourches, Judy A. Strickland, Nicole Kleinstreuer, Carolina Horta Andrade · 2014
Repetitive exposure to a chemical agent can induce an immune reaction in susceptible individuals leading to skin sensitization. We have developed computational models capable of accurately assessing the skin sensitization potential of environmental chemicals. To this end, we have (i) compiled, curated, and integrated the largest publicly-available database of skin-sensitizing chemicals; (ii) used this data to generate and validate QSAR models for skin sensitization; and (iii) employed these models to identify putative sensitizers among chemicals in the Scorecard and Tox21 databases. A random forest method was employed for QSAR modeling of compounds characterized by SiRMS and Dragon descriptors, and the OECD compliant model validation workflow was followed. The overall classification accuracies of QSAR models discriminating sensitizers from non-sensitizers were 68-88% when evaluated on several external validation sets. When compared to the OECD QSAR toolbox skin sensitization module, our models afforded significantly higher Positive and Negative Predictive Rates. When applied to chemicals within the applicability domains, the models could reliably identify positive and negative sensitizers with 94% and 71% certainty, respectively. Statistically significant descriptors from high-accuracy models yielded SAR rules that could guide structural optimization of chemicals of interest. Using these models, we have identified putative skin sensitizers in the ScoreCard and Tox21 databases as primary hits for further experimental testing. The addition of ToxCast in vitro assays as biological descriptors for hybrid QSAR modeling improved the quality of resulting models and enabled deeper interpretation of these models in terms of underling MOAs.