QSAR Models for Oral Bioavailability and Volume of Distribution and Their Application to the Mapping of the TK Space of Endocrine Disruptors

Toxicokinetics (TK) properties are essential in the framework of chemical risk assessment and drug discovery. Indeed, a TK profile provides information about the fate of chemicals in the human body. In this context, Quantitative Structure-Activity Relationship (QSAR) models are convenient computational tools to predict TK properties. Here, we developed QSAR models for the prediction of two TK properties: oral bioavailability and volume of distribution at steady state (VDss). We collected and curated two large sets of 1712 and 1591 chemicals for respectively oral bioavailability and VDss and compared regression and classification (binary and multiclass) models with the application of several machine learning algorithms. The best predictive performance of the models for regression (R) prediction is characterized by a Q²F3 equal to 0.34 with the R-CatBoost model, for oral bioavailability, and a Geometric Mean Fold Error (GMFE) equal to 2.35 for VDss with the R-RF model. The models were then applied on a list of potential endocrine disrupting chemicals (EDC), highlighting chemicals with a high probability of posing a concern on human health due to their TK profile. Based on the results obtained, insights about structural determinants of TK properties for EDCs are also further discussed.