Robust Moiety Model Selection Using Mass Spectrometry Measured Isotopologues
Stable isotope resolved metabolomics (SIRM) experiments uses stable isotope tracers to provide superior mass spectroscopy (MS) and nuclear magnetic resonance (NMR) metabolomics datasets for metabolic flux analysis and metabolic modeling. Several software packages exist for metabolic flux analysis when provided a metabolic model and appropriate isotopomer and/or isotopologue datasets, mostly from 13C tracer time series experiments. However, assumptions of model correctness can seriously compromise interpretation of metabolic flux results generated from these packages. Therefore, we have developed a metabolic modeling software package specifically designed for moiety model comparison and selection based on the metabolomics data provided. This moiety modeling framework facilitates analysis of time-series SIRM MS isotopologue profiles using a set of plausible moiety models and data in a JSONized representation. The moiety_modeling Python package is available on GitHub and the Python Package Index and provides facilities for model parameter optimization, analysis of optimization results, and model selection. Furthermore, this package is capable of analyzing multi-tracer datasets. Here, we tested the effectiveness of this moiety modeling framework in model selection with two sets of time-series MS isotopologue datasets for uridine diphosphate N-acetyl-D-glucosamine (UDP-GlcNAc) generated from different MS platforms: direct infusion nanoelectrospray Fourier transform MS and liquid chromatography MS. Results generated from the analyses of these datasets demonstrate the robustness of our model selection methods by the successful selection of the optimal model from over 40 models provided. Also, the effects of specific optimization methods, degree of optimization, selection criteria, and specific objective functions on model selection are illustrated. Furthermore, different types of error can exist in the datasets, and proper selection of the objective function can help reduce the optimization side effects caused by the specific types of uncertainty in these datasets. Overall, these results indicate that over-optimization can lead to failure in model selection, but combining multiple datasets can help prevent this overfitting effect. The implication is that SIRM datasets in public repositories of reasonable quality can be combined with newly acquired datasets to improve model selection. Furthermore, curation efforts of public metabolomics repositories to maintain high data quality could have huge impacts on future metabolic modeling efforts.