PFAS急性経口毒性におけるリードクロス構造解析：Isalos Analytics Platformの自動機械学習を活用

The ubiquity and environmental persistence of per- and polyfluoroalkyl substances (PFASs) have raised significant concerns about their detrimental effects on human health. Collective scientific efforts are increasingly focused on elucidating PFAS toxicity mechanisms and identifying potential low-impact PFAS structures that retain the exceptional properties of this chemical class. To advance the use of in silico methods in PFAS toxicity assessment, we developed a robust modelling framework for predicting PFAS acute oral toxicity class (high or low) in rats, leveraging the enhanced capabilities of the in-house Isalos Analytics Platform. The automated machine learning (autoML) functionality was employed to optimise four ML models-k-nearest neighbours (kNN), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and fully connected neural network (NN)-using Mold2 molecular descriptors, and to identify the top-performing model through five-fold cross-validation. The selected kNN model (k = 3) was used for predictions on the held-out testing set, achieving an accuracy of 81.5%, while a Shapley values analysis provided valuable insights into the factors influencing toxicity predictions. Furthermore, the nearest-neighbour-based methodology enabled a read-across structural analysis of PFAS similarity groups consisting of each testing set instance and its three closest neighbours in the training set. This analysis revealed a consistent association between polyaromatic and heterocyclic structural features and high acute oral toxicity. The developed, thoroughly validated read-across model is freely accessible through the INSIGHT RatTox web application as well as the INSIGHT Cheminformatics Platform in Enalos Cloud, supporting high-throughput screening of PFAS compounds and investigation of structural similarities with their nearest neighbours for enriched structural interpretation.