• Developed SkinCast, an AOP-based interpretable AI model integrating KE1–KE4 predictions. • Biological KE weighting enabled transparent, mechanistic-based adverse outcome prediction. • A GCN architecture achieved balanced high performance and outperformed existing QSAR tools. • Screening 3,415 fragrance ingredients identified 15 previously unclassified sensitizers. Consumer product ingredients released into the environment may cause secondary human exposure and potential adverse outcomes (AOs) such as skin sensitization, a complex immunological process defined by four key events (KEs) within the AO pathway. To enhance interpretability and predictive reliability, we developed SkinCast, an AI-driven, mechanistically interpretable model that integrates these KEs to estimate the final AO. Multiple machine learning and deep learning algorithms were evaluated for each KE, and a graph convolutional network was selected for its superior structural representation. A biologically weighted scoring function linked KE-level predictions to the AO based on their biological relevance, enabling mechanistic outcome estimation. SkinCast exhibited robust predictive performance (area under the receiver operating characteristic curve = 0.81–0.90 across KE1–KE4), high concordance with human patch test data, and outperformed established quantitative structure–activity relationship (QSAR) tools. Applicability domain analysis showed that 90.9% of validation compounds were within the reliable chemical space. When applied to 3,415 fragrance ingredients representing environmentally released consumer product ingredients, SkinCast identified 15 potential sensitizers lacking existing GHS classifications. This approach provides an interpretable predictive model for identifying potential skin sensitizers arising from environmentally released consumer product ingredients.
Lee et al. (Sun,) studied this question.