Tetramethylammonium hydroxide (TMAH) and substances that release tetramethylammonium (TMA) are classified as Priority Assessment Chemical Substances (PACSs) under registration number 17 of the Japan Chemical Substances Control Law (CSCL, 1973). This classification requires a thorough human health hazard assessment and derivation of Hazard Assessment Value (HAVs) for the oral and inhalation exposure at the Assessment II stage. We analyzed their general, developmental, reproductive toxicity, genotoxicity, and carcinogenicity using hazard data from both domestic and international risk assessment agencies and subsequently proposed an HAV. For oral exposure, a no-observed-adverse-effect-level (NOAEL) of 1 mg/kg/day, based on transient or persistent salivation in parent rats from a TMAH developmental and reproductive toxicology (DART) screening study, was chosen as the point of departure (POD). The POD was then divided by uncertainty factors (UFs) totaling 1,000 (interspecies variation: 10, intraspecies variation: 10, short study duration: 10), resulting in an oral HAV of 0.001 mg/kg/day for TMAH. Due to a lack of hazard data for humans and animals via inhalation, an HAV for the inhalation route was not established.

Drug-induced phospholipidosis (DIPL) is linked to various toxicities, including hepatotoxicity, making it a critical screening factor in the early stages of drug discovery. Several models based on chemical structures have been constructed to predict compounds with DIPL potential. However, most of these models only classify results as inducers or non-inducers, without identifying the specific substructures responsible for positive outcomes. To address this limitation, we constructed an artificial intelligence (AI) model to predict compounds with DIPL potential and visualize structural alerts. The proposed model was constructed using kMoL, an open-source software library that employs a graph neural network approach to learn from chemical structural data. We employed the bagging method, resulting in a model with a high predictive performance. The model attained an F1 score of 0.796 on the external test set. In addition, we used the integrated gradient method to visualize the substructures that contributed to positive predictions. When applying the method to compounds that experimentally conducted structure-activity relationship investigations, the proposed AI model accurately predicted DIPL potential, demonstrating its practical utility in early-stage drug discovery. By predicting DIPL based on the chemical structure of compounds, the proposed model can aid in the screening for DIPL, potentially improving the safety profile of new drug candidates.