Unraveling the Linkages between Molecular Abundance and Stable Carbon Isotope Ratio in Dissolved Organic Matter Using Machine Learning

Dissolved organic matter (DOM) is a complex mixture of
molecules
that constitutes one of the largest reservoirs of organic matter on
Earth. While stable carbon isotope values (δ13C)
provide valuable insights into DOM transformations from land to ocean,
it remains unclear how individual molecules respond to changes in
DOM properties such as δ13C. To address this, we
employed Fourier transform ion cyclotron resonance mass spectrometry
(FT-ICR MS) to characterize the molecular composition of DOM in 510
samples from the China Coastal Environments, with 320 samples having
δ13C measurements. Utilizing a machine learning model
based on 5199 molecular formulas, we predicted δ13C values with a mean absolute error (MAE) of 0.30‰ on the
training data set, surpassing traditional linear regression methods
(MAE 0.85‰). Our findings suggest that degradation processes,
microbial activities, and primary production regulate DOM from rivers
to the ocean continuum. Additionally, the machine learning model accurately
predicted δ13C values in samples without known δ13C values and in other published data sets, reflecting the
δ13C trend along the land to ocean continuum. This
study demonstrates the potential of machine learning to capture the
complex relationships between DOM composition and bulk parameters,
particularly with larger learning data sets and increasing molecular
research in the future.