Prediction of NDMA formation potential using non-target analysis data: a proof of concept

N-Nitrosodimethylamine (NDMA) is a nitrogenous disinfection by-product (DBP) that has been included in drinking water regulations worldwide because of its carcinogenicity and hazardousness. Anticipating the NDMA formation potential (FP) of a water sample before its disinfection is a complex task, since the formation of this DBP is promoted by an overwhelmingly long and heterogeneous list of miscellaneous precursors. In the present study, we explored different predictive models, based on high-resolution mass spectrometry (HRMS) non-target data, to accurately estimate the NDMA-FP of complex environmental waters. The samples included tertiary effluents and wastewater-impacted river waters, all of which were taken in the frame of a short-term full-scale water reclamation trial. Non-target analysis, conducted by liquid chromatography (LC) coupled to (Orbitrap) HRMS, provided an extensive dataset with 3924 unknown molecular features. The peak list was curated and refined with the criteria ubiquity, sensitivity, intensity, and orthogonality in order to obtain a reduced list of 42 robust and independent variables. The occurrence of known NDMA precursors could not explain satisfactorily the relatively high NDMA-FP of the samples and its variability (85 ± 13–840 ± 3 ngNDMA l−1). In contrast, simple linear models built with non-target HPLC-HRMS data were able to predict the NDMA-FP values with normalised root-mean-square deviations (NRMSDs) of ∼11–15% after model training and cross-validation. These results were improved by regression decision trees (8.1 ± 4.2% NRMSD) and k-nearest neighbour classification models (Matthews correlation coefficient >0.9). Overall, our results indicate that non-target data, in combination with predictive analytics, have a great potential to estimate the NDMA-FP of actual environmental samples, which opens the door to its application in water treatment management and DBP control.