Perfluorocarboxylic acid toxicity analysis with albumin/1-anilinonaphthalene-8-sulfonate probe assisted by machine learning

Perfluorocarboxylic acids (PFCAs) exist as mixtures in aquatic environments. However, routine PFCA monitoring remains instrument-intensive and often reports concentrations rather than mixture toxicity for water risk management. Herein, we developed a machine-learning-assisted fluorescence framework that estimates the perfluorooctanoic acid (PFOA)-equivalent toxicity ( PFOAe ) for aqueous PFCA mixtures. The method used a simple bovine serum albumin/1-anilinonaphthalene-8-sulfonate (BSA/ANS) probe. The machine learning framework was developed by utilizing a stacked ensemble model, including XGBoost and CatBoost base-learners with a Ridge meta-learner to decipher the nonlinear relationships between excitation emission matrices and PFOAe . On the log 10 ( PFOAe ) scale, the model achieved a root mean squared error (RMSE) of 0.286 on a held-out test set, and an RMSE of 0.280 for spiked tap-water samples. Five authentic electroplating wastewater samples further provided a limited real sample feasibility validation. The limit of detection was 0.126 µmol l -1 PFOAe (52.2 µg l -1 as PFOA), enabling rapid screening and prioritization of moderately-to-highly contaminated waters rather than trace-level analysis. Moreover, the reconciliation between chemical fluorescence phenomena and data-driven mechanistic interpretation suggests that the model behavior is consistent with the underlying chemical mechanism. Our method provides a lower complexity and risk-relevant screening tool to prioritize water samples for further confirmatory instrumental analysis and support monitoring-based risk management for PFCA mixtures.