The applicability domain’s definition is important in validating the reliability of toxicity prediction models. However, applicability domains are established in various model formats and situations, and the definition method is still in the research stage.

In this study, because predicting the correctness of a binary classifier means correctly identifying the target on which the binary classifier operates, we considered that the correctness prediction model itself functions as a threshold for the applicability domain of the binary classifier and used machine learning to develop in silico models that define applicability domains by predicting the true/false values of the ITSv2-defined approach for identifying skin sensitization hazards of chemical substances. The correctness of the ITSv2 hazard identification was used as the objective variable, and each alternative to animal testing data, molecular descriptors, and similar substances’ information were used as an explanatory variable. Because very few data were misclassified, they were very difficult to detect.

Therefore, as a different approach, we assumed the distribution of the false data and quantitatively predicted the misclassification risk of the qualitative evaluation. By setting thresholds for the misclassification risk, the true/false values of the ITSv2 hazard identification were judged. With the newly defined applicability domain, the accuracy of the ITSv2 hazard identification improved substantially. In this study, machine learning was used to optimize the method for defining applicability domains according to trends derived from toxicity assessment data.