Consensus model for skin sensitization assessment using a rule-based model and LLNA and GPMT statistics-based models

• In silico tools assess sensitization via key event/adverse outcome pathways. • A new consensus model combines three independent key event/adverse outcome pathways for assessment. • Tested on chemicals, the model achieved a balanced accuracy of 78%. • The model covers diverse chemical spaces with high prediction accuracy. The potential for skin sensitization has traditionally been assessed in vivo ; however, animal welfare concerns, the trend toward restrictions, and the prohibition of the use of animals have led to a shift toward the use of non-animal alternatives such as in vitro and in silico tools. In silico tools mainly include rule-based and statistics-based models. Although the use of multiple computational methods is recommended, many tools consist of only one method. Furthermore, skin sensitization develops through multiple key event (KE)/adverse outcome (AO) pathways, but many in silico tools consist of only one KE/AO. We constructed a consensus model based on three different independent skin sensitization KE/AOs from a rule-based model, a local lymph node assay (LLNA) statistics-based model, and a guinea pig maximization test (GPMT) statistics-based model. The rule-based model is based on KE1 and considers the metabolism of pre- and pro-haptens. The LLNA and GPMT statistics-based models are based on KE4 and AO, respectively, and characterized by the use of approximately 2000 and 3000 chemicals in the training dataset, respectively. These models use larger datasets than those previously reported. The constructed consensus model was tested on chemicals labeled with human results from OECD Guideline 497. The results showed that the performance of the majority-voting model was the highest, with a balanced accuracy of 78%. The model combines a wide range of chemical spaces with high prediction accuracy.