Deep Learning Based Metabolite Annotation.

Area of Science:

• Analytical Chemistry
• Computational Biology
• Biochemistry

Background:

• Metabolite annotation is a critical challenge in untargeted metabolomics using liquid chromatography-mass spectrometry (LC-MS).
• Limited public spectral libraries containing tandem mass spectrometry (MS/MS) data hinder accurate identification of unknown compounds.
• Machine learning offers a promising avenue to predict molecular properties from MS/MS spectra.

Purpose of the Study:

• To improve the accuracy of molecular fingerprint prediction from MS/MS data.
• To enhance metabolite identification in LC-MS-based untargeted metabolomics.
• To investigate the utility of high-dimensional spectral and fingerprint representations.

Main Methods:

• Utilized a convolutional neural network (CNN) for molecular fingerprint prediction.
• Employed MS/MS spectral data from the MoNA repository and NIST 20.
• Explored high-dimensional representations of spectral data and molecular fingerprints.

Main Results:

• The study investigated methods to improve accuracy in molecular fingerprint prediction using MS/MS spectra.
• High-dimensional representations were explored to enhance prediction performance.
• The approach aids in ranking candidate metabolite IDs for spectra not found in existing libraries.

Conclusions:

• Machine learning, particularly CNNs, can effectively predict molecular fingerprints from MS/MS data.
• This predictive capability assists in annotating metabolites lacking spectral library matches.
• Investigating high-dimensional data representations shows potential for further improving metabolomics annotation accuracy.