Structure-Based Prediction and Ranking of PFAS Groundwater Mobility for EU Drinking Water Regulation

We present a machine learning framework that simultaneously predicts the three physicochemical properties governing PFAS subsurface mobilitysoil sorption (log KOC, R² = 0.90), water solubility (log Sw, R² = 0.90), and ionization class (93% accuracy)from molecular structure alone.

The models were trained on 122 PFAS spanning 37 subclasses using 44 molecular descriptors, including nine that encode the segmentation of the fluorinated backbone by ether linkages, and were validated by cross-validation, external test sets, leave-one-subclass-out analysis, Y-randomization, and a formal applicability domain. SHAP analysis indicates that ether-linkage descriptors are associated with lower predicted soil retention, providing a model-based explanation for the high mobility of GenX-type replacements.