Predicting Bioconcentration Factors of Per- and Polyfluoroalkyl Substances Using a Directed Message Passing Neural Network with Multimodal Feature Fusion

Amid growing concerns regarding the ecological risks posed by emerging contaminants, per- and polyfluoroalkyl substances (PFASs) present significant challenges for risk assessment due to their structural diversity and the paucity of experimental data on their bioaccumulation. This study investigated the bioconcentration factors (BCFs) of 18 emerging and legacy PFASs using zebrafish in a flow-through exposure system and constructed a robust BCF prediction model to address the data gaps associated with numerous novel PFASs. Experimental results indicated that perfluoro(3,5,7,9,11-pentaoxadodecanoic) acid (PFO5DoDA) and perfluoro-2,5-dimethyl-3,6-dioxanonanoic acid (C9 HFPO-TA) exhibited higher bioaccumulation potential than perfluorooctanoic acid (PFOA). A multimodal feature-fused directed message passing neural network (FF-DMPNN) model was constructed, integrating molecular graph representations, physicochemical descriptors, and bioassay data reflecting absorption, distribution, metabolism, and excretion characteristics. The FF-DMPNN model demonstrated superior predictive performance compared to conventional machine learning approaches by providing a more complete representation of molecular structures and physicochemical properties, achieving higher accuracy (R² = 0.742) and robustness in predicting BCF values for PFASs. Application of the model to a comprehensive PFAS database identified 2.45% of chemicals as bioaccumulative, highlighting the need for regulatory attention. Overall, this study provides critical insights into the bioconcentration risks associated with PFASs and offers a reliable framework for prioritizing regulatory actions for these emerging contaminants, addressing a pressing need for their effective environmental management.