IDSL_MINT: a deep learning framework to predict molecular fingerprints from mass spectra.

Area of Science:

• Metabolomics and Exposomics
• Computational Chemistry
• Bioinformatics

Background:

• Untargeted metabolomics and exposomics studies generate vast amounts of unannotated tandem mass spectrometry (MS/MS) spectra, hindering biological interpretation.
• Structural annotation of MS/MS spectra is a significant bottleneck in advancing biological insights from these datasets.

Purpose of the Study:

• To introduce IDSL_MINT, a deep learning framework designed to translate MS/MS spectra into molecular fingerprint descriptors.
• To enable users to train customizable models using their own MS/MS spectral libraries and molecular fingerprints.
• To improve the annotation rates of MS/MS spectra in untargeted metabolomics and exposomics studies.

Main Methods:

• Developed IDSL_MINT, a deep learning framework leveraging transformer models for mass spectrometry data analysis.
• Trained models on user-provided reference MS/MS libraries and customizable molecular fingerprint descriptors.
• Benchmarked IDSL_MINT using the LipidMaps database and a test study dataset.

Main Results:

• IDSL_MINT successfully translated MS/MS spectra into molecular fingerprint descriptors.
• The framework demonstrated improved annotation rates for previously unannotated MS/MS spectra when benchmarked against existing spectral libraries.
• Performance was validated using the LipidMaps database.

Conclusions:

• IDSL_MINT offers a powerful and customizable solution for spectral annotation in metabolomics and exposomics.
• The framework has the potential to significantly increase the overall annotation rates in large-scale untargeted studies.
• IDSL_MINT is available via GitHub, promoting accessibility and further development.