Work

MSThunder: A deep learning-based nontargeted analytical framework for the accurate and rapid identification of unknown organic pollutants in water

Year: 2025

Type: article

Abstract: Nontargeted analysis (NTA) has emerged as a crucial technology for identifying unknown organic pollutants in water. However, current NTA methods commonly omit or misjudge the pollutants due to the low-accuracy and efficiency in the structural prediction of mass spectrometry. Here, we developed a novel NTA framework named MSThunder for rapid and accurate identification of unknown organic pollutants in water. We first constructed a method for obtaining product ion formulas with isotopes to expand additional features into the input while mitigates noise interference from the matrix. Then, we established a transformer model that transitions from formula to structural sequences, less constrained by the lack of training data. MSThunder achieved 59.1 % top-1 accuracy in the benchmark (CASMI 2016 Challenge), outperforming other major existing models. The highest 77.1 % of chemical structures across all real-world experimental samples spiked with a trace mixture of 96 pollutants were correctly inferred, superior to both SIRIUS (74.0 %) and TraceFinder (46.9 %). The highest prediction speed of 0.32 s per spectrum was achieved among existing models. Furthermore, 181 pollutants were identified in various water samples of Yellow River Basin by MSThunder. Three pollutants with high potential hazards, including triallyl cyanurate, harmaline and 5,7-dimethyl-3-methylsulfonyl-[1,2]thiazolo[3,4- d ]pyrimidine-4,6-dione, were first found in the water samples. This will promote the screening of unknown hazardous organic pollutants in water and provide the fundamental methods for water quality safety assessment.

Source: Environmental Chemistry and Ecotoxicology

DOI: 10.1016/j.enceco.2025.07.022