Identifying priority PBT-like compounds from emerging PFAS by nontargeted analysis and machine learning models

Abstract

As traditional per and polyfluoroalkyl substances (PFAS) are phased out, emerging PFAS are being developed and widely used. However, little is known about their properties, including persistence, bioaccumulation, and toxicity (PBT). Screening for emerging PFAS relies on available chemical inventory databases. Here, we compiled a database of emerging PFAS obtained from nontargeted analysis and assessed their PBT properties using machine learning models, including qualitative graph attention networks, Insubria PBT Index and quantitative EAS-E Suite, VEGA, and ProTox-II platforms. Totally 282 homologues (21.8% of emerging PFAS) were identified as PBT based on the combined qualitative and quantitative prediction, in which 140 homologues were detected in industrial and nonbiological/biological samples, belong to four categories, i.e. modifications of perfluoroalkyl carboxylic acids, perfluoroalkane sulfonamido substances, fluorotelomers and modifications of perfluoroalkyl sulfonic acids. Approximately 10.1% of prioritized emerging PFAS were matched to chemical vendors and 19.6% to patents. Aqueous film-forming foams and fluorochemical factories are the predominant sources for prioritized emerging PFAS. The database and screening results can update the assessment related to legislative bodies such as the US Toxic Substances Control Act and the Stockholm Convention. The combined qualitative and quantitative machine learning models can provide a methodological tool for prioritizing other emerging organic contaminants.

Keywords: Emerging PFAS; Machine learning; Nontargeted analysis; Prioritized list; Screening.