Machine learning-powered pseudo-target screening of emerging contaminants in water: A case study on tetracyclines

Identifying emerging contaminants (ECs) in complex water environment is one of the greatest challenges. Target screening (TS) is limited by the lack of reference standards, whereas non-target screening (NTS) is subject to complex and unreliable data processing. In this study, we reported the machine learning (ML)-powered pseudo-target screening (PTS) for primary identification of ECs with tetracyclines (TCs) serving as model. Based on mass spectrometry (MS) data collected from MassBank database, we performed data purification by removing interferential peaks through optimizing the threshold factor (P=1%), the parameter that reflected intensity of interferential peaks (A) in relative to that of maximum peak (A_max). Then, the well-trained XGBoost model was obtained for correct identification of TCs and Non-TCs with probability approaching 100% by feeding experimental MS data with integrated peak- and test-related features. We for the first time demonstrated the effectiveness of such feature integration strategy for improving accuracy, reliability and anti-interference ability offered by the ML models. The XGBoost model could also identify the TCs that were in the both presence and absence of model training set, suggesting potential generalizability for identifying the unregulated and unknown ECs. Compared with previously reported TS and NTS, our ML-powered PTS framework offered an efficient, simple and reliable alternative to identifying ECs in environmental samples without the need for prior knowledge. This study not only has important implications for dealing with accidental emergency of water pollution relevant to occurrence of ECs, but also represents paradigm shift to develop AI-powered algorithm frameworks for identifying more ECs beyond tested TCs herein.