In Search of Disentanglement in Tandem Mass Spectrometry Datasets.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Tandem mass spectrometry (MS/MS) is crucial for metabolite identification.
• Current MS/MS data analysis faces challenges with instrument variability and database querying.
• Developing interpretable and instrument-agnostic representations is essential for advancing metabolomics.

Purpose of the Study:

• To apply generative modeling and representation learning to MS/MS spectra.
• To investigate the disentanglement of MS/MS spectra into underlying generation factors.
• To create interpretable and instrument-agnostic digital representations of metabolites.

Main Methods:

• Utilized variational autoencoders (VAEs) for generative modeling and representation learning.
• Applied VAEs to tandem mass spectrometry data with minimal prior knowledge of generation factors.
• Developed a two-step approach for selecting disentangled VAE models.

Main Results:

• Demonstrated that VAEs can effectively disentangle MS/MS spectra data.
• Identified meaningful latent representations aligned with known factors of variation (e.g., collision energy, ionization mode).
• Showcased the potential for instrument-agnostic data analysis and improved metabolite identification.

Conclusions:

• Generative modeling with VAEs offers a powerful approach for analyzing MS/MS spectra.
• Disentangled representations facilitate instrument-agnostic comparisons and database queries.
• The developed methodology can be extended to other high-dimensional and complex datasets.