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

Area of Science:

• Environmental Chemistry
• Analytical Chemistry
• Toxicology

Background:

• High-resolution mass spectrometry (HRMS) is vital for identifying per- and polyfluoroalkyl substances (PFAS).
• Accurate quantification of non-targeted PFAS in complex matrices like vegetables is challenging due to a lack of standards and matrix interference.
• This hinders environmental fate and human exposure assessments.

Purpose of the Study:

• To develop a machine learning combined with internal standards (ML-IS) model for accurate PFAS quantification in vegetable matrices using HRMS.
• To improve the accuracy of non-targeted PFAS concentration predictions.

Main Methods:

• Established a machine learning combined with internal standards (ML-IS) model for HRMS.
• Applied the model to quantify PFAS in Shanghai cabbage and white radish.
• Compared ML-IS model performance against conventional methods and previous ML models.

Main Results:

• The ML-IS model demonstrated excellent predictive performance, improving accuracy 1.5- to 8.7-fold compared to models without internal standards.
• The ML-IS model significantly outperformed semi-quantitative and previous ML methods.
• The model maintained performance across different HRMS platforms and was used to quantify six suspect PFAS in real samples.

Conclusions:

• The ML-IS model advances HRMS capabilities for robust, quantitative predictions of unknown PFAS in complex vegetable matrices.
• This approach is crucial for accurate environmental monitoring and risk assessment of PFAS contamination.