Leveraging latent information in NMR spectra for robust predictive models

Area of Science:

• Metabolomics
• Nuclear Magnetic Resonance (NMR) Spectroscopy
• Chemometrics

Background:

• Extracting biologically meaningful insights from complex 1D NMR spectra is a key challenge in metabolomics.
• Existing data representation techniques may limit model interpretability and robustness.

Purpose of the Study:

• To compare the effectiveness of "Spectral Binning" and "Targeted Profiling" for representing 1D NMR spectra in metabolomics.
• To evaluate the impact of variable scaling techniques on predictive model performance.

Main Methods:

• Simulated 1D NMR spectra with defined characteristics were used to build multivariate statistical models.
• Two data representation techniques, Spectral Binning and Targeted Profiling, were assessed.
• The influence of different variable scaling methods was analyzed.

Main Results:

• Targeted Profiling resulted in more interpretable predictive models than Spectral Binning.
• Targeted Profiling demonstrated greater robustness against compound overlap and variations in solution conditions (pH, ionic strength).
• Findings from simulated data were confirmed with a real-world dataset.

Conclusions:

• Targeted Profiling is a superior method for data representation in 1D NMR metabolomics compared to Spectral Binning.
• This technique enhances model interpretability and reliability, facilitating more accurate biological data extraction.