Abstract
【Introduction】Skin sensitization is important toxicity for the safety assessment of chemicals. Currently, it is need to develop new safety evaluation methods that don’t use animals. We aimed to develop a regression model to predict skin sensitization intensity (EC3) by using toxicity database and machine learning. Moreover, to improve interpretability, we interpreted prediction results using SHAP values. 【Method】We used 154 substances from Guideline (OECD No. 497, 2021) as model building and internal validation, and 72 substances from document (Hoffmann, 2022) as external validation. The regression model was built using XGBoost with EC3 as the objective variable. In vitro tests on adverse outcome pathway (AOP), chemicals information were used as explanatory variables. The applicability domain was set based on the k-nearest neighbor method and in vitro test results. 【Result and Discussion】We calculated the percentage of substances whose predicted values were within 1/5 to 5 times true values within the applicability domain to evaluate the performance of the model. As a result, we found 80% for internal validation and 65% for external validation were within the range. Based on SHAP values, it was confirmed that in vitro test results contributed significantly and positively to the prediction, suggesting that this model takes AOP into account. Using toxicity information and machine learning, we were able to construct a model to predict EC3. In addition, SHAP values allowed us to check the contribution to prediction and to build a model with interpretability.
