Abstract
Environmental pollution monitoring represents a major challenge due to the growing presence of a large and diverse number of potential contaminants. In complement to target analysis, nontarget analysis, via liquid chromatography (LC) coupled to high-resolution mass spectrometry (HRMS), is increasingly used to provide a more comprehensive characterization of pollution. The challenge associated with this type of analysis is particularly related to the data treatment for substance identification. One of the main limitations is that all data must be manually reviewed, which is tedious and time-consuming. Machine learning algorithms aim to reproduce human behavior, and their capabilities were therefore evaluated to automatically identify substances in suspect screening approaches. After selecting the relevant features produced from LC/HRMS, seven different machine learning models were evaluated for each of the three different databases, which resulted in the selection of logistic regression (LR) and random forest (RF)-based algorithms. An interface was built to rank the identified substances and to assess the performance of the developed models. The LR model provided the best results when retention times were available. The developed LR and RF models were determined complementarily, particularly when no retention times were available. However, limitations were noticed when using a database containing different HRMS technologies.