JESTR: Joint Embedding Space Technique for Ranking Candidate Molecules for the Annotation of Untargeted Metabolomics...

Area of Science:

• Metabolomics
• Computational chemistry
• Bioinformatics

Background:

• Metabolomics annotation, the process of identifying molecular structures from mass spectral data, faces challenges with low accuracy rates.
• Existing methods for molecule-to-spectra and spectra-to-molecular fingerprint prediction have limitations in improving annotation efficiency.

Purpose of the Study:

• To introduce JESTR, a novel computational paradigm for enhanced metabolomics annotation.
• To address the limitations of current annotation techniques by leveraging a joint embedding space for molecules and spectra.

Main Methods:

• JESTR embeds molecular structures and mass spectra into a shared latent space.
• Candidate molecule structures are ranked against query spectra using cosine similarity in the joint space.
• The model incorporates regularization with candidate molecules during training to refine performance.

Main Results:

• JESTR demonstrates superior performance in metabolomics annotation compared to existing tools, achieving an average improvement of 23.6% - 71.6% for rank@[1-5] across three datasets.
• Regularization during training significantly boosted rank@1 performance by 11.4% and improved the model's ability to differentiate between correct and incorrect molecular assignments.
• The JESTR approach offers a promising new avenue for accurate metabolite identification.

Conclusions:

• JESTR represents a significant advancement in metabolomics annotation, offering higher accuracy and improved discernment of molecular structures.
• The joint embedding space paradigm provides a powerful framework for future developments in spectral data analysis.
• This work facilitates deeper insights into the metabolome by enabling more reliable identification of metabolites.