Learning from All Views: A Multiview Contrastive Framework for Metabolite Annotation.

Area of Science:

• Metabolomics
• Computational Chemistry
• Bioinformatics

Background:

• Metabolomics, utilizing high-throughput mass spectrometry, is crucial for understanding cellular biochemistry, disease mechanisms, drug development, and personalized medicine.
• Low spectral annotation rates due to the challenge of assigning molecular structures to measured spectra hinder advancements in metabolomics.

Purpose of the Study:

• To introduce MultiView Projection (MVP), a novel framework designed to improve spectral annotation rates in metabolomics.
• To create a joint embedding space for molecules and spectra by integrating multiple data views.

Main Methods:

• MVP employs contrastive multiview learning to capture mutual information across diverse data views: molecular graphs, molecular fingerprints, spectra, and consensus spectra.
• The framework learns from all views jointly, unlike prior methods that used concatenation or auxiliary tasks.
• MVP supports flexible annotation using either individual spectra or consensus spectra.

Main Results:

• MVP significantly improves molecular candidate ranking for spectral annotation.
• Annotation using query consensus spectra with MVP outperforms rank aggregation strategies based on constituent spectrum annotation.
• MVP achieves 36.0% rank@1 for mass-based retrieval and 14.0% for formula-based retrieval using consensus spectra.

Conclusions:

• MVP provides a flexible and extensible foundation for learning from multiple molecule/spectra data views.
• The framework demonstrates superior or comparable performance to existing methods for spectral annotation.
• Improved annotation rates facilitated by MVP can accelerate discoveries in metabolomics-related fields.