A deep learning architecture for combining and imputing heterogeneous metabolomics datasets

Area of Science:

• Metabolomics
• Bioinformatics
• Computational Biology

Background:

• Public metabolomics databases contain numerous datasets valuable for understanding disease mechanisms.
• Existing datasets are often sparse, measuring only a fraction of metabolites, hindering machine learning model development.
• Combining sparse datasets directly results in high dimensionality and missing values, posing challenges for analysis.

Purpose of the Study:

• To develop and evaluate novel methods for merging sparse metabolomics datasets.
• To improve the imputation of missing metabolite data for enhanced machine learning model training.
• To effectively combine diverse metabolomics datasets while minimizing data gaps.

Main Methods:

• Proposing two novel dataset merging approaches: Iterative similarity-based merging and Model-guided agglomerative merging.
• Ensuring a minimum sparsity threshold is maintained during dataset merging.
• Employing Variational Autoencoders (VAE) for imputation model training on merged datasets.
• Evaluating the methods on the complete set of public datasets from the Metabolomics Workbench.

Main Results:

• The proposed merging and imputation methods significantly outperform the state-of-the-art approach.
• Achieved substantially improved imputation performance on sparse metabolomics data.
• Demonstrated the effectiveness of combining diverse datasets while managing metabolite overlap.

Conclusions:

• The novel merging strategies effectively address sparsity in public metabolomics datasets.
• VAE-based imputation on merged datasets enhances the ability to analyze complex molecular mechanisms in diseases.
• These methods provide a robust framework for leveraging large-scale metabolomics data for biomedical research.