IodoFinder: Machine Learning-Guided Recognition of Iodinated Chemicals in Nontargeted LC-MS/MS Analysis

Iodinated
disinfection byproducts (I-DBPs) pose significant health
concerns due to their high toxicity. Current approaches to recognize
unknown I-DBPs in mass spectrometry (MS) analysis rely on negative
ionization mode, in which the characteristic I– fragment
can be observed in tandem mass spectra (MS/MS). Still, many I-DBPs
ionize exclusively in positive ionization mode, where the I– fragment is absent. To address this gap, this work developed a machine
learning-based strategy to recognize iodinated compounds (I-compounds)
from their MS/MS in both electrospray positive (ESI+) and negative
ionization (ESI−) modes. Investigating over 6000 MS/MS spectra
of 381 I-compounds, we first identified five characteristic I-containing
neutral losses and one diagnostic I– fragment in
ESI+ and ESI– modes, respectively. We then trained Random Forest
models and integrated them into IodoFinder, a Python program, to streamline
the recognition of I-compounds from raw LC-MS data. IodoFinder accurately
recognized over 96% of the 161 I-compound standards in both ionization
modes. In its application to DBP mixtures, IodoFinder discovered 19
I-DBPs with annotated structures and an additional 17 with assigned
formulas, including 12 novel and 3 confirmed I-DBPs. We envision that
IodoFinder will advance the identification of both known and unknown
I-compounds in exposome studies.