Machine Learning Tools Can Pinpoint High-Risk Water Pollutants ^†

Liquid chromatography–high-resolution mass spectrometry (LC/HRMS) is a powerful tool for detecting chemicals that are present in low concentrations. While this technique has revealed thousands of ionizable pollutants in environmental samples [1,2], the expanding list of emerging contaminants highlights the urgency to speed up their risk assessment [3,4]. Generally, the risk assessment workflow starts with structural identification, followed by obtaining the analytical standard for confirmation, toxicity assessment, and quantification with a calibration curve. To speed up the process, machine learning has found use in predicting toxicity and ionization efficiency; however, most of the models in use still require a chemical structure as an input. Therefore, detected but unidentified chemicals are frequently discarded from further analysis and the bioactivity of samples often remains partially unexplained [5]. Still, the fragmentation spectrum provides information about the structure which can be related to the properties of the chemical. We developed a workflow for estimating the risk of chemicals detected in non-target screening based on their MS² data. Two prediction models, MS2Quant [6] for ionization efficiency and MS2Tox [7] for acute fish toxicity, were trained based on structural fingerprints. While structural fingerprints can be calculated from a structure, the recently developed SIRIUS+CSI:FingerID software [8] offers the possibility to predict these fingerprints based on the MS² spectrum, and therefore predict chemical properties without structural assignment. Based on the validation set, the root mean square errors of MS2Quant and MS2Tox were 5.9× (39 chemicals) and 7.8× (219 chemicals), respectively. These models were applied in a non-target screening workflow regarding wastewater analysis. The preliminary results show that MS2Quant and MS2Tox help to pinpoint chemicals that pose a higher risk compared to a top five approach. Therefore, this approach provides the possibility to evaluate the risk of unidentified LC/HRMS features and prioritize high-risk chemicals in identification.