Advancing Fluorine Mass Balance Using Machine Learning-Based Quantitative Nontarget Screening
4 affiliations
- 1Department of Plant and Environmental Science, University of Copenhagen, Thorvaldsensvej 40, Frederiksberg C, Denmark
- 2Department of Environmental Science (ACES), Stockholm University, Svante Arrheniusväg 8, Stockholm, Sweden
- 3Luxembourg Centre for Systems Biomedicine (LCSB)
- 4University of Luxembourg
Abstract
Per- and polyfluoroalkyl substances (PFAS) often account for only a small fraction of the extractable organic fluorine (EOF) in fluorine mass balance (FMB) studies. Closing this fluorine gap requires identification and quantification of a broader range of fluorinated compounds, including low-fluorinated substances that are difficult to prioritize by nontarget screening (NTS). Here, a workflow combining fluorine-specific NTS with machine learning-based quantification (qNTS) to improve FMB closure was developed. Extreme Gradient Boosting (XGBoost)-based qNTS models for electrospray ionization (ESI) + and ESI − were trained with 597 reference compounds (185 fluorinated) reaching mean absolute error of ×2.8 and ×2.5 on held-out folds, respectively. An independent validation set of 57 compounds previously unseen by the model reached a mean error of ×2.1. The workflow was applied to urban runoff extracts and AFFF-impacted soils. In sewage-associated runoff, fluorinated pharmaceuticals and their transformation products dominated the identified fluorine, whereas conventional PFAS (i.e. those containing a perfluoroalkyl chain) contributed <3%. Non-PFAS organofluorines (including many pharmaceuticals) contributed up to 48% of the EOF. Identified compounds explained 54–107% of measured EOF in sewage-associated samples but only 1–24% in street runoff, indicating a large fraction of unidentified fluorine in the latter. Application to aqueous film forming foam (AFFF)-impacted soils showed that qNTS performed well for anionic PFAS but was less reliable for cationic and zwitterionic PFAS, which were poorly represented in the training data. Overall, qNTS provided concentration estimates that enabled rapid identification of the compounds contributing most to EOF, thereby directing subsequent target confirmation and FMB closure.
Related
- Processing-and cell-stage thermodynamic screening of nacre-and bone-inspired carbonate and phosphate phases on aluminium current collectors in fluorine-free lithium-ion cells with an open screening framework
- Advancing specific local wine identification through integration of ICP-MS mineral wine profiles and Machine Learning: A case study on discrimination among the ten Beaujolais crus
- Radiolabeling with fluorine-18 of a protein, interleukin-1 receptor antagonist
- 氟功能化锆基金属有机框架负载薄膜微萃取阵列用于草甘膦残留的高通量分析 Fluorine-functionalized zirconium-based metal-organic framework immobilized thin film microextraction array for high-throughput analysis of glyphosate residues
- Lewis Base-Induced Valence Tautomerism in a Phosphorus Cation Enables the Activation of Strong Element-Fluorine Bonds
BibTeX
@misc{zweigle2026advancing,
title = {{Advancing Fluorine Mass Balance Using Machine Learning-Based Quantitative Nontarget Screening}},
author = {Jonathan Zweigle and Merle Plassmann and Thomas M. Karlsson and Emma L. Schymanski and Jonathan P. Benskin and Jan H. Christensen},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008377/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008377/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Advancing Fluorine Mass Balance Using Machine Learning-Based Quantitative Nontarget Screening
AU - Jonathan Zweigle
AU - Merle Plassmann
AU - Thomas M. Karlsson
AU - Emma L. Schymanski
AU - Jonathan P. Benskin
AU - Jan H. Christensen
PY - 2026
DA - 2026/09/06
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008377/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008377/v1
AB - Per- and polyfluoroalkyl substances (PFAS) often account for only a small fraction of the extractable organic fluorine (EOF) in fluorine mass balance (FMB) studies. Closing this fluorine gap requires identification and quantification of a broader range of fluorinated compounds, including low-fluorinated substances that are difficult to prioritize by nontarget screening (NTS). Here, a workflow combining fluorine-specific NTS with machine learning-based quantification (qNTS) to improve FMB closure was developed. Extreme Gradient Boosting (XGBoost)-based qNTS models for electrospray ionization (ESI) + and ESI − were trained with 597 reference compounds (185 fluorinated) reaching mean absolute error of ×2.8 and ×2.5 on held-out folds, respectively. An independent validation set of 57 compounds previously unseen by the model reached a mean error of ×2.1. The workflow was applied to urban runoff extracts and AFFF-impacted soils. In sewage-associated runoff, fluorinated pharmaceuticals and their transformation products dominated the identified fluorine, whereas conventional PFAS (i.e. those containing a perfluoroalkyl chain) contributed <3%. Non-PFAS organofluorines (including many pharmaceuticals) contributed up to 48% of the EOF. Identified compounds explained 54–107% of measured EOF in sewage-associated samples but only 1–24% in street runoff, indicating a large fraction of unidentified fluorine in the latter. Application to aqueous film forming foam (AFFF)-impacted soils showed that qNTS performed well for anionic PFAS but was less reliable for cationic and zwitterionic PFAS, which were poorly represented in the training data. Overall, qNTS provided concentration estimates that enabled rapid identification of the compounds contributing most to EOF, thereby directing subsequent target confirmation and FMB closure.
ER -