DeePFAS: Deep-Learning-Enabled Rapid Annotation of PFAS: Enhancing Nontargeted Screening through Spectral Encoding and Latent Space Analysis.

Abstract

Detecting PFAS is challenging due to their diverse chemical structures, lack of standards, complex sample matrices, and the need for sensitive equipment to measure trace levels. Background contamination and the sheer number of PFAS further hinder the development of a universal detection method. Liquid chromatography-high-resolution mass spectrometry (LC-HRMS) is the primary tool capable of analyzing PFAS in water, soil, and biological samples, and it is widely adopted in regulatory testing. However, LC-HRMS faces challenges, including contamination risk, labor-intensive preparation, low detection limits, and time-consuming data processing that requires advanced software and expertise to distinguish structurally similar compounds. To overcome these obstacles, we present DeePFAS, a deep-learning-based method for rapid annotation of PFAS. DeePFAS employs a spectral encoder integrating convolutional and transformer architectures to project raw MS2 spectra into a latent space of chemical structural features learned from a large corpus of unlabeled compounds. DeePFAS enables efficient annotation of MS2 spectra by comparing latent representations with candidate molecules, streamlining large-scale nontargeted PFAS screening, and reducing analytical complexity. Our method demonstrates the potential of AI-driven tools in environmental chemistry and is available at https://github.com/CMDM-Lab/DeePFAS.