Longitudinal Metabolomics Data Analysis Informed by Mechanistic Models

Area of Science:

• Metabolomics
• Systems Biology
• Computational Biology

Background:

• Metabolomics data often suffers from noise, small sample sizes, and missing values, hindering accurate analysis.
• While data-driven methods are useful, incorporating prior knowledge of metabolic pathways can significantly improve insights.
• Existing methods may not fully leverage mechanistic understanding for metabolomics data interpretation.

Purpose of the Study:

• To introduce a novel data analysis approach for metabolomics that integrates mechanistic models.
• To enhance the analysis of noisy and incomplete metabolomics data by combining real measurements with simulated data.
• To improve the discovery of biologically relevant patterns and biomarkers.

Main Methods:

• Time-resolved metabolomics data from plasma samples (COPSAC2000 cohort) were structured as a third-order tensor (subjects x metabolites x time).
• Simulated data from a human whole-body metabolic model were also structured as a tensor (virtual subjects x metabolites x time).
• Coupled tensor factorizations were employed to jointly analyze real and simulated data, coupled in the metabolite mode.

Main Results:

• Joint analysis of real and simulated data showed improved pattern discovery and higher correlation with BMI-related phenotypes in males compared to analyzing real data alone.
• Performance was comparable between joint and real-data-only analysis in females.
• The approach demonstrated robustness in handling incomplete measurements but highlighted limitations with incorrect prior information.

Conclusions:

• Joint analysis using coupled tensor factorizations effectively integrates mechanistic prior information into metabolomics data analysis.
• This hybrid approach guides the interpretation of real data and reveals more interpretable patterns.
• The method offers a promising strategy for robust metabolomics analysis, particularly when dealing with data imperfections.