Variational autoencoders learn transferrable representations of metabolomics data.

Area of Science:

• Metabolomics
• Systems Biology
• Computational Biology

Background:

• High-dimensional metabolomics data analysis often relies on dimensionality reduction.
• Current methods struggle to identify nonlinear relationships within metabolomics datasets.
• Variational Autoencoders (VAEs) offer a deep learning approach to learn nonlinear latent representations.

Purpose of the Study:

• To apply VAEs for deconvolution of complex metabolomics data.
• To investigate the biological interpretability of VAE-derived latent spaces.
• To assess the generalizability of VAEs on unseen clinical metabolomics datasets.

Main Methods:

• Training a VAE on a large-scale human blood metabolomics cohort (>4500 individuals).
• Utilizing a global feature importance score to analyze latent space pathway composition.
• Evaluating VAE performance against linear and nonlinear Principal Component Analysis (PCA).

Main Results:

• Latent dimensions learned by the VAE represent distinct cellular processes.
• VAE successfully generated meaningful latent representations for unseen datasets (type 2 diabetes, AML, schizophrenia).
• VAE representations showed stronger correlations with clinical groups than PCA-derived dimensions.

Conclusions:

• VAEs are effective in capturing nonlinearities in metabolomics data.
• The VAE model learns biologically relevant and transferable latent representations.
• This deep learning method enhances the analysis of complex metabolic data for clinical applications.