Interpretable Machine Learning and Reactomics Assisted Isotopically Labeled FT-ICR-MS for Exploring the Reactivity and Transformation of Natural Organic Matter during Ultraviolet Photolysis

Abstract

Isotopically labeled FT-ICR-MS combined with multiple post-analyses, including interpretable machine learning (IML) and a paired mass distance (PMD) network, was employed to unravel the reactivity and transformation of natural organic matter (NOM) during ultraviolet (UV) irradiation. FT-ICR-MS analysis was used to assign formulas, which were classified on the basis of their molecular compositions and structural categories. Isotope (deuterium, D) labeling was utilized to unequivocally determine the photochemical products and examine the development of OD radical-mediated NOM transformation. With regard to the reactive molecular formulas, CHOS formulas exhibited the highest reactivity (86.5% of precursors disappeared) followed by CHON (53.4%) and CHO (24.6%) formulas. With regard to structural categories, the degree of reactivity decreased in the following order: tannins > condensed aromatics > lignin/CRAMs. The IML algorithm demonstrated that the crucial features governing the reactivity of formulas were the molecular weight, DBE-O, NOSC, and the presence of heteroatoms (i.e., N and S), suggesting that the large and unsaturated compounds containing S and N are more prone to photodegradation. The reactomics approach using the PMD network further indicated that 11 specific molecular formulas in the CHOS and CHO class served as hubs, implying a higher photoreactivity and participation in a range of transformations. The isotope labeling analyses also found that, among the reactions observed, hydroxylation (i.e., +OD) is dominant for lignin/CRAMs and condensed aromatics, and formulas containing ≤10 D atoms were developed. Overall, this study, by adopting rigorous and interpretable techniques, could provide in-depth insights into the molecular-level dynamics of NOM under UV irradiation.