FLARE: Fine-grained Learning for Alignment of spectra-molecule REpresentation Enhances Metabolite Annotation.

Area of Science:

• Chemistry
• Computational Biology
• Biochemistry

Background:

• Metabolomics relies on accurate metabolite annotation, which is challenging with tandem mass spectrometry.
• Current implicit models align spectra and structures in embedding space but miss detailed peak-substructure relationships.

Purpose of the Study:

• Introduce FLARE (Fine-grained Learning for Alignment of spectra-molecule REpresentations) for improved metabolite annotation.
• Leverage bidirectional peak-node alignment for chemically meaningful local correspondences.

Main Methods:

• Developed a contrastive learning framework (FLARE).
• Employed bidirectional peak-node alignment under learned weak supervision.
• Computed similarity via maxima over peak-to-atom and atom-to-peak interactions.

Main Results:

• Achieved state-of-the-art performance on MassSpecGym (43.15% rank@1 mass-based, 22.66% formula-based).
• Surpassed previous models by over 63% in accuracy.
• Demonstrated FLARE's learned embeddings correlate with molecular classes and fingerprint similarity.

Conclusions:

• FLARE enables interpretable spectra-molecule matching by capturing local chemical relationships.
• Successfully detected differential metabolites in a breast cancer xenograft study, highlighting translational potential.