Evaluation and application of machine learning-based retention time prediction for suspect screening of pesticides and pesticide transformation products in LC-HRMS

Computational QSAR models have gradually been preferred for retention time prediction in data mining of emerging environmental contaminants using liquid chromatography coupled with mass spectrometry. Generally, the model performance relies on the components such as machine learning algorithms, chemical features, and example data. In this study, we evaluated the performances of four algorithms on three feature sets, using 321 and 77 pesticides as the training and validation sets, respectively. The results were varied with different combinations of algorithms on distinct feature sets. Two strategies including enhancing the complexity of chemical features and enlarging the size of the training set were proved to improve the results. XGBoost, Random Forest, and lightGBM algorithms exhibited the best results when built on a large-scale chemical descriptors, while the Keras algorithm preferred fingerprints. These four models have comparable prediction accuracies that at least 90% of pesticides in validation set can be successfully predicted with ΔRT <1.0 min. Meanwhile, a blended prediction strategy using average results from four models presented a better result than any single model. This strategy was used for assisting identification of pesticides and pesticide transformation products in 120 strawberry samples from a national survey of food contamination. Twenty pesticides and twelve pesticide transformation products were tentatively identified, where all pesticides and two pesticide transformation products (bifenazate diazene and spirotetramat-enol) were confirmed by standard materials. The outcome of this study suggested that retention time prediction is a valuable approach in compound identification when integrated with in silico MS2 spectra and other MS identification strategies.