Development of an in silico evaluation system that quantitatively predicts skin sensitization using OECD Guideline No. 497 ITSv2 defined approach for skin sensitization classification.

The Integrated Testing Strategy version 2 (ITSv2) Defined Approach, which is a reliable skin sensitization hazard and multi-step risk assessment method, does not support quantitative risk assessment such as local lymph node assay EC3 values. In this study, we developed a high-performance in silico evaluation system that quantitatively predicts the EC3 values of chemical substances by combining the ITSv2 Defined Approach for hazard identification (ITSv2 HI) with machine learning models. This system uses in chemico/in vitro test data, molecular descriptors, and distance information based on read-across concepts as explanatory variables. The system achieves an R2 value of 0.617 on external-validation data. Substances misclassified in ITSv2 HI are considered to have properties that do not match the correspondence between tests expressing the adverse outcome pathway assumed in the ITSv2 Defined Approach and skin sensitization. Therefore, ITSv2 HI is assumed to be correct within the applicability domains of this system. When using only substances within the applicability domains to reconstruct CatBoost models, the R2 value reached 0.824 on the external-validation data, representing an improvement in system performance. The results demonstrate the utility of explanatory variables that reflect the read-across concept and the advantages of integrating multiple prediction methods.

Chemical characterization of anemia-inducing aniline-related substances and their application to the construction of a decision tree-based anemia prediction model.

Anemia is a well-observed toxicity of chemical substances, and aniline is a typical anemia-inducing substance. However, it remains unclear whether all aniline-like substances with various substituents could induce anemia. We thus investigated the physicochemical characteristics of anemia-inducing substances by decision tree analyses. Training and validation substances were selected from a publicly available database of rat repeated-dose toxicity studies, and discrimination models were constructed by decision tree and bootstrapping methods with molecular descriptors as explanatory variables. To improve the accuracy of discrimination, we individually evaluated the explanatory variables to modify them, established "prerules" that were applied before subjecting a substance to a decision tree by considering metabolism, such as azo reduction and N-dealkylation, and introduced the idea of "partly negative" evaluation for substances having multiple aniline-like substructures. The final model obtained showed 79.2% and 77.5% accuracy for the training and validation dataset, respectively. In addition, we identified some chemical properties that reduce the anemia inducibility of aniline-like substances, including the addition of a sulfonate or carboxy functional group and/or a bulky multiring structure to anilines. In conclusion, the present findings will provide a novel insight into the mechanistic understanding of chemically induced anemia and help to develop a prediction system.