Machine learning-based in silico quantification framework for non-targeted PFAS in complex vegetable matrices

High-resolution mass spectrometry (HRMS) is a powerful tool for the comprehensive identification of per- and polyfluoroalkyl substances (PFAS). However, the lack of reference standards and interference from complex matrices (such as vegetables) pose significant challenges for accurate quantification of non-targeted PFAS, hindering the assessment of their environmental fate and exposure to humans. Therefore, the present study established a machine learning combined with internal standards (ML-IS) model on HRMS instruments to achieve accurate quantification of PFAS in Shanghai cabbage (Brassica chinensis) and white radish (Raphanus sativus) matrices. The optimized ML-IS model achieved excellent predictive performance for non-targeted PFAS concentrations, representing a 1.5- to 8.7-fold improvement in prediction accuracy compared to the model without internal standards. Furthermore, the proposed ML-IS model substantially outperformed both the conventional semi-quantitative method based on structural similarity and the previously published ML model, while also maintaining satisfactory predictive performance across PFAS datasets from other HRMS platforms. Moreover, the ML-IS model was successfully applied to quantify six suspect PFAS in real vegetable samples. These results mark an important step towards advancing HRMS from a tool for qualitative detection to one capable of robust concentration predictions for unknown PFAS in complex vegetable matrices.