Enhancing pharmaceutical hazard assessment with machine learning mobility models

Assessing the environmental mobility of pharmaceuticals is critical for robust chemical hazard assessment, yet experimental sorption data are often limited. Furthermore, existing predictive models for sorption often lack reliability, particularly for ionisable compounds and frequently neglect the effects of sorbent properties on mobility. Here, we present an optimised and accessible random forest machine learning model that predicts linear sorption coefficients (log K_d) for active pharmaceutical ingredients, capturing soil-dependent variance. Model outputs show good agreement with external literature data (R² = 0.60), and moderate performance on industry data within the training domain (R² = 0.29). To demonstrate utility, predicted log K_d values for 1661 human and veterinary pharmaceuticals were normalised to organic carbon content (log K_oc) across seven theoretical soil types to classify their mobility against EU regulatory criteria. In total, 228 pharmaceuticals were classified as very mobile and 671 as mobile. Comparison with empirical classifications showed good agreement (63.2% agreement with industry data, 52.4% for literature data), suggesting that our optimised model provides a valuable tool to support the hazard and risk assessment of pharmaceuticals and, potentially, other classes of organic micropollutants.