Electrospray Ionization Efficiency Predictions and Analytical Standard Free Quantification for SFC/ESI/HRMS

• Published in: Journal of the American Society for Mass Spectrometry 2023-07, Vol.34 (7), p.1511-1518
• Main Authors: Bieber, Stefan, Letzel, Thomas, Kruve, Anneli
• Format: Article
• Language: English
• Subjects: ionization efficiency; machine learning; nontarget screening; quantification; supercritical fluid chromatography
• ISSN: 1044-0305; 1879-1123; 1879-1123
• DOI: 10.1021/jasms.3c00156

Supercritical fluid chromatography (SFC) is a promising, sustainable, and complementary alternative to liquid chromatography (LC) and has often been coupled with high resolution mass spectrometry (HRMS) for nontarget screening (NTS). Recent developments in predicting the ionization efficiency for LC/ESI/HRMS have enabled quantification of chemicals detected in NTS even if the analytical standards of the detected and tentatively identified chemicals are unavailable. This poses the question of whether analytical standard free quantification can also be applied in SFC/ES/HRMS. We evaluate both the possibility to transfer an ionization efficiency predictions model, previously trained on LC/ESI/HRMS data, to SFC/ESI/HRMS as well as training a new predictive model on SFC/ESI/HRMS data for 127 chemicals. The response factors of these chemicals ranged over 4 orders of magnitude in spite of a postcolumn makeup flow, expectedly enhancing the ionization of the analytes. The ionization efficiency values were predicted based on a random forest regression model from PaDEL descriptors and predicted values showed statistically significant correlation with the measured response factors (p < 0.05) with Spearman’s rho of 0.584 and 0.669 for SFC and LC data, respectively. Moreover, the most significant descriptors showed similarities independent of the chromatography used for collecting the training data. We also investigated the possibility to quantify the detected chemicals based on predicted ionization efficiency values. The model trained on SFC data showed very high prediction accuracy with median prediction error of 2.20×, while the model pretrained on LC/ESI/HRMS data yielded median prediction error of 5.11×. This is expected, as the training and test data for SFC/ESI/HRMS have been collected on the same instrument with the same chromatography. Still, the correlation observed between response factors measured with SFC/ESI/HRMS and predicted with a model trained on LC data hints that more abundant LC/ESI/HRMS data prove useful in understanding and predicting the ionization behavior in SFC/ESI/HRMS.