DeePFAS, a novel deep-learning model, streamlines large-scale non-targeted screening of "forever chemicals" (PFAS) by projecting raw MS2 spectra into a latent space of chemical features, offering a rapid, AI-driven solution to replace complex traditional analysis.
The accurate detection of per- and polyfluoroalkyl substances (PFAS), often termed "forever chemicals," presents a critical and complex challenge for environmental science due to their structural diversity, the lack of standardized methods, and the need for highly sensitive equipment to measure trace environmental levels.
A study published in Environmental Science and Technology reveals an innovative, deep learning-based approach to overcome these obstacles. The prevalence of background contamination and the sheer number of distinct PFAS compounds further complicate the development of universal detection protocols.
Current analytical methodology relies primarily on Liquid Chromatography–High-Resolution Mass Spectrometry (LC-HRMS), which is widely adopted for analyzing PFAS in various matrices (water, soil, biological samples, etc.).
However, LC-HRMS introduces significant challenges, including high contamination risk, labor-intensive sample preparation, and time-consuming data processing. This processing demands advanced software and expertise, particularly to distinguish structurally similar compounds.
To address these limitations, a team of researchers at National Taiwan University introduced DeePFAS, a novel deep-learning method for the rapid and efficient annotation of PFAS compounds.
DeePFAS utilizes a specialized spectral encoder, which integrates convolutional and transformer architectures, to translate raw MS2 spectra (chemical fingerprints) into a "latent space." This latent space represents a concise mapping of chemical structural features learned from a large corpus of unlabeled compounds.
By comparing these latent representations with candidate molecules, DeePFAS enables highly efficient annotation of MS2 spectra. This approach significantly streamlines large-scale nontargeted PFAS screening efforts and reduces the overall analytical complexity of environmental monitoring.
The study successfully demonstrated DeePFAS's sensitivity in identifying PFAS-specific features, and its practical feasibility was confirmed through application to real-world wastewater samples.
The team acknowledged minor limitations, including occasional false-positive annotations associated with certain compounds where the tool exhibited lower confidence.
Future work will involve expanding the chemical fingerprint library using in silico methods and comparing DeePFAS results with those from existing tools to boost confidence. Ultimately, DeePFAS offers a robust, open-source AI solution for researchers and is available for further research and development on GitHub.
"This approach is designed to enhance nontargeted PFAS analysis and significantly reduce analytical complexity," says Prof. Yufeng Jane Tseng, corresponding author of the study.
To see article on Phys.org: https://phys.org/news/2025-10-deepfas-ai-tool-advances-chemical.html