Optimal machine learning algorithm for prediction model for retention times of plant toxins

In suspect screening or nontargeted analysis via LC high-resolution MS (LC-HRMS), high-accuracy identification typically relies on retention times (RTs) and MS–MS spectra. However, RTs are difficult to obtain due to the scarcity of reference standards. Here, we developed a Quantitative Structure Retention Relationships (QSRR) -based RT prediction model using machine learning, specifically for plant toxins implicated in accidental food poisoning. A dataset for QSRR model development was generated using the molecular descriptors (MDs) and experimental RTs of 524 compounds. QSRR models were constructed as regression models derived from the relationship between experimental RTs and MDs using 10 machine learning algorithms. The QSRR model with support vector regression (SVR) outperformed the other QSRR models in generalization on the analyzed dataset (R²: 0.972, mean absolute error: 183 [approximately 1.6 min], mean absolute percentage error [MAPE]: 6%; Q²: 0.875, MAE: 584 [approximately 2.0 min], MAPE: 15%). Furthermore, the SVR QSRR model successfully predicted the RTs of nine plant toxins with errors of ±0.5 min. Thus, this model enhances the confidence level of plant toxin identification via LC-HRMS.