Systematic classification of unknown metabolites using high-resolution fragmentation mass spectra

Area of Science:

• Computational chemistry
• Metabolomics
• Bioinformatics

Background:

• Nontargeted tandem mass spectrometry identifies thousands of molecules.
• Structural molecule annotation is limited by available libraries and databases.
• This restricts the analysis and interpretation of experimental metabolomics data.

Purpose of the Study:

• To develop a computational tool for systematic compound class annotation.
• To enable the analysis of molecules lacking spectral or structural reference data.
• To predict compound classes lacking tandem mass spectrometry training data.

Main Methods:

• Developed CANOPUS (class assignment and ontology prediction using mass spectrometry), a deep neural network tool.
• Trained the network to predict 2,497 compound classes from fragmentation spectra.
• Evaluated performance using reference data and compared against four baseline methods.

Main Results:

• CANOPUS achieved very high prediction performance with an average accuracy of 99.7% in cross-validation.
• The tool outperformed four baseline methods in compound class prediction.
• Demonstrated broad utility in diverse biological and chemical investigations.

Conclusions:

• CANOPUS enables systematic compound class annotation for previously unidentifiable molecules.
• The tool significantly enhances metabolomics analysis and interpretation.
• Provides biological insights at the compound class level across various applications.