Modifying Chromatography Conditions for Improved Unknown Feature Identification in Untargeted Metabolomics

Untargeted metabolomics is an essential
component of systems biology
research, but it is plagued by a high proportion of detectable features
not identified with a chemical structure. Liquid chromatography–tandem
mass spectrometry (LC–MS/MS) experiments produce spectra that
can be searched against databases to help identify or classify these
unknowns, but many features do not generate spectra of sufficient
quality to enable successful annotation. Here, we explore alterations
to gradient length, mass loading, and rolling precursor ion exclusion
parameters for reversed phase liquid chromatography (RPLC) and hydrophilic
interaction liquid chromatography (HILIC) that improve compound identification
performance for human plasma samples. A manual review of spectral
matches from the HILIC data set was used to determine reasonable thresholds
for search score and other metrics to enable semi-automated MS/MS
data analysis. Compared to typical LC–MS/MS conditions, methods
adapted for compound identification increased the total number of
unique metabolites that could be matched to a spectral database from
214 to 2052. Following data alignment, 68.0% of newly identified features
from the modified conditions could be detected and quantitated using
a routine 20-min LC–MS run. Finally, a localized machine learning
model was developed to classify the remaining unknowns and select
a subset that shared spectral characteristics with successfully identified
features. A total of 576 and 749 unidentified features in the HILIC
and RPLC data sets were classified by the model as high-priority unknowns
or higher-importance targets for follow-up analysis. Overall, our
study presents a simple strategy to more deeply annotate untargeted
metabolomics data for a modest additional investment of time and sample.