Host: The Japanese Society of Toxicology
Name : The 50th Annual Meeting of the Japanese Society of Toxicology
Date : June 19, 2023 - June 21, 2023
As one of the alternative methods to animal tests for evaluating the safety of chemical substances, in silico methods have attracted a great deal of attention due to recent advances in machine learning technology. ITSv2 is a reliable skin sensitization hazard and multi-step risk assessment method listed in OECD Guideline No. 497, but does not support quantitative risk assessment such as LLNA EC3 values. In this study, we developed an in silico evaluation system that quantitatively predicts EC3 of chemical substances by combining ITSv2 hazard identification and machine learning models. This system used in chemico/in vitro test data, molecular descriptors and distance information based on read-across concepts. Performance of the prediction system was 0.862 for training data, 0.550 for internal-validation data, and 0.617 for external-validation data as R2 value. Moreover, substances that were misclassified in ITSv2 hazard identification are considered to have properties that don't match the correspondence between tests expressing the AOP assumed in ITSv2 and skin sensitization, therefore, the correctness of ITSv2 hazard identification was determined as the threshold of the applicability limitation of this prediction system. When the performance was re-evaluated by excluding the substances misclassified in ITSv2 hazard identification, R2 value was 0.963, 0.681 and 0.815 for each data. In addition, by reconstructing the CatBoost models using only substances within the applicability limitation, 0.995, 0.787 and 0.824 for each data, further improving the prediction performance.
