Enhanced Environmental PFAS Characterization Using a Virtual High-Resolution Mass Spectral Library Generated by...

Per- and polyfluoroalkyl substances (PFAS) represent a critical class of persistent environmental contaminants with significant ecological and human health implications. However, the rapid emergence of novel PFAS has far outpaced the development of reference mass spectral databases. Here, Neural Per- and Polyfluoroalkyl Substances Mass Spectrometry (NPFAS-MS), a transfer learning-based neural network model, was developed to predict PFAS-specific high-resolution mass spectra. NPFAS-MS was fine-tuned from a pretrained model using PFAS tandem mass (MS/MS) spectra. NPFAS-MS outperformed other in silico spectral prediction models for PFAS spectra prediction across multiple spectral similarity metrics. In library searching tasks, libraries generated by other spectral prediction models showed top-1 recall between 42.1% and 55.4%, while NPFAS-MS demonstrated 71.1%. Applying the virtual PFAS mass spectral library generated with NPFAS-MS using 10,553 PFAS structures from the U.S. EPA and NORMAN databases to groundwater and aqueous film-forming foam (AFFF) samples revealed more potential PFAS than other mass spectral databases. Specifically, 38 potential PFAS were annotated in AFFF products and 40 in groundwater samples. NPFAS-MS enabled characterization of emerging PFAS, including ultrashort-chain, unsaturated, and substituted derivatives in environmental matrices. This advancement enables comprehensive environmental monitoring of rapidly evolving PFAS contamination. NPFAS-MS and associated resources were deployed as a web-based tool at https://cosbi10.ee.ncku.edu.tw/NPFAS_MS/, enabling both structure-to-spectrum prediction and library searching against 31,659 predicted PFAS spectra.