Leveraging latent information in NMR spectra for robust predictive models

A significant challenge in metabolomics experiments is extracting biologically meaningful data from complex spectral information. In this paper we compare two techniques for representing 1D NMR spectra: "Spectral Binning" and "Targeted Profiling". We use simulated 1D NMR spectra with specific characteristics to assess the quality of predictive multivariate statistical models built using both data representations. We also assess the effect of different variable scaling techniques on the two data representations. We demonstrate that models built using Targeted Profiling are not only more interpretable than Spectral Binning models, but are more robust with respect to compound overlap, and variability in solution conditions (such as pH and ionic strength). Our findings from the synthetic dataset were validated using a real-world dataset.