Do experimental projection methods outcompete retention time prediction models in non-target screening? A case study...

Area of Science:

• Analytical Chemistry
• Environmental Science
• Mass Spectrometry

Background:

• Retention time (RT) is critical for structure elucidation in non-target screening (NTS) using liquid chromatography–high-resolution mass spectrometry (LC-HRMS).
• Existing methods for RT estimation include projection and prediction approaches, which may face challenges due to differences between source/training chromatographic systems (CS_source/CS_training) and the NTS system (CS_NTS).

Purpose of the Study:

• To evaluate the generalizability of RT projection and prediction models across diverse chromatographic systems (CSs) commonly used in NTS.
• To identify key chromatographic parameters influencing the accuracy of these RT estimation methods.

Main Methods:

• Utilized data from a NORMAN interlaboratory comparison involving 41 calibration chemicals and 45 suspects analyzed across 37 different CSs.
• Assessed the performance of both RT projection and machine learning (ML) prediction models against experimental RTs.

Main Results:

• The accuracy of both projection and prediction models was directly correlated with the similarity between the chromatographic systems.
• Mobile phase pH and column chemistry were identified as the most significant factors impacting RT estimation accuracy.
• Prediction models performed comparably to projection models when CS_training was similar to CS_NTS, even if CS_source differed significantly.

Conclusions:

• The generalizability of RT estimation models is highly dependent on chromatographic system similarity, particularly mobile phase pH and column chemistry.
• Machine learning models for RT prediction should incorporate mobile phase and column chemistry parameters for improved accuracy.
• The choice between projection and prediction models should consider the similarity between training/source and NTS chromatographic systems.