Development and Application of Machine Learning Models for Predicting Dermal Absorption with Consideration of Vehicle Effects

The accurate evaluation of human dermal absorption is critical for the risk assessment of topically applied chemicals. While in vitro testing is common, in silico models, as part of New Approach Methodologies (NAMs), have gained attention for simplicity and efficiency. Accurate prediction requires considering the skin permeability coefficient ( K p ) and diffusion coefficient ( D ), as well as the influence of vehicles. In this study, a database was constructed using human in vitro dermal absorption data, comprising 639 log K p and 135 log D values. A comparative analysis of machine learning algorithms identified Random Forest (RF) as the most optimal approach. For the log K p model, the molecular weight (MW c ), log P of chemicals (log P c ), log P of vehicles (log P v ), topological polar surface area (TPSA), hydrogen bond acceptor 1 (HBA1) and Henry's law constant (H) were selected as descriptors. For the log D model, MW c , log P c and TPSA were selected. These models predicted dermal absorption rates (DA (%)) for 25 external chemicals, showing almost no underestimation compared to in vitro values. Overall, 56% of the chemicals were predicted within a ten-fold range of the experimental data, while 44% were overpredicted by more than ten-fold. Their applicability for photosafety exposure assessment of benzophenone derivatives (BZPs) was examined, with predicted K p rankings similar to published data. In conclusion, the proposed in silico models effectively predicted DA (%) and may contribute to exposure assessment, with potential applicability to photosafety risk assessment.