An integrated deep learning framework for the interpretation of untargeted metabolomics data

Area of Science:

• Biochemistry
• Systems Biology
• Computational Biology

Background:

• Untargeted metabolomics is crucial for understanding biological mechanisms and clinical outcomes.
• A significant challenge is the uncertainty in matching experimental features to known metabolites due to experimental limitations and complex biological processes.
• Existing methods struggle with one-to-many feature-metabolite relationships, impacting the reliability of metabolite selection and pathway analysis.

Purpose of the Study:

• To develop an integrated deep learning framework to address matching uncertainty in metabolomics data analysis.
• To simultaneously evaluate metabolite importance, infer feature-metabolite matching likelihood, and select disease-relevant metabolic subnetworks.
• To provide a robust method for uncovering biological mechanisms from complex metabolomics datasets.

Main Methods:

• An integrated deep learning framework utilizing a gradual sparsification neural network.
• Incorporation of known metabolic networks and feature-metabolite annotation relationships into the model architecture.
• Simultaneous optimization for metabolite importance, feature-metabolite matching, and disease subnetwork selection.

Main Results:

• The deep learning framework effectively handles matching uncertainty in metabolomics data.
• The model successfully identified interpretable metabolic subnetworks in COVID-19 and aging mouse brain datasets.
• Achieved simultaneous evaluation of metabolite importance, feature-metabolite matching likelihood, and disease subnetwork selection.

Conclusions:

• The proposed deep learning framework offers a robust solution for metabolomics data analysis, particularly in addressing feature-metabolite matching uncertainty.
• This approach enhances the reliability of metabolite selection and biological interpretation, facilitating the discovery of critical metabolic pathways.
• The method demonstrates broad applicability, as shown by its successful application to diverse datasets including those from COVID-19 research and aging studies.