Integrating machine learning, suspect and nontarget screening reveal the interpretable fates of micropollutants and their transformation products in sludge

Activated sludge enriches vast amounts of micropollutants (MPs) when wastewater is treated, posing potential environmental risks. While standard methods typically focus on target analysis of known compounds, the identity, structure, and concentration of transformation products (TPs) of MPs remain less understood. Here, we employed a novel approach that integrates machine learning for the quantification of nontarget TPs with advanced target, suspect, and nontarget screening strategies. 39 parent chemicals and 286 TPs were identified, with the majority being pharmaceuticals, followed by phthalate acid ester and alkylphenols. To quantify TPs without reference standards, we applied machine learning to forecast the relative response factors (RRFs) relied on their physicochemical characteristics. The random forest regression model showed great performance, with prediction errors of RRFs ranging from 0.03 to 0.35. The mean concentrations for parents and TPs were 1.32 -19.83 and 6.35 -9.94 μg/g dw, respectively. Further risk-based prioritization integrating environmental exposure and ToxPi scoring ranked the identified 182 compounds, with three parents and one TP recognized as high priorities for management. N-demethylation and N-oxidated TPs are generally less toxic than their parents. These findings are expected to facilitate MPs and their TPs investigations for reliable environmental monitoring and risk assessment across different sludge treatment processes.