In-Depth Analysis of the Data from an Interlaboratory Study of Quantitative Non-Target Screening—How Do the Instrumental Methods Compare?

TL;DR

This study compares different methods for quantifying contaminants in environmental samples and finds that machine learning approaches outperform traditional methods, though some factors affect results.

Contribution

The study provides insights into the limitations and variability of instrumental methods for predicting ionization efficiency in non-target screening.

Findings

Prediction errors were not linked to specific instrument parameters but were influenced by organic modifiers and additives.

Comparable logRFs were observed across datasets when using a linear model, though variability was higher for compounds with low logRF.

The developed dashboard allows free access to the data for further analysis.

Abstract

Non-target screening utilizing liquid chromatography–high-resolution mass spectrometry is increasingly employed for the environmental monitoring of contaminants; however, obtaining quantitative results of detected suspected compounds is challenging; different approaches have been suggested. A recent interlaboratory comparison of quantification approaches showed that the machine learning-based approaches leveraging predicted ionization efficiencies outcompete surrogate standard-based approaches, independent of the method used. In this study, we further analyzed data from the interlaboratory comparison to: (1) evaluate whether the prediction errors could be linked to instrument parameters; and (2) investigate the comparability of response factors (RFs) across different datasets to shed light on the limitations of the instrumental method on the predicted ionization efficiency approach. No specific parameters could be linked to systematic effects on the prediction errors; however, the choice of organic modifier and/or additive type influenced the detection of some compounds. Comparable logRFs across datasets were observed when a linear model was used to project the values to the same scale. Nevertheless, the projected logRF scale was compressed for datasets with low similarity to the anchoring dataset. Moreover, compounds with low logRF showed higher variability across the datasets. The data are freely available and can be interrogated in the developed dashboard.

Figures