ChemEmbed: a deep learning framework for metabolite identification using enhanced MS/MS data and multidimensional...

Area of Science:

• Computational chemistry
• Metabolomics
• Bioinformatics

Background:

• Mass spectrometry (MS/MS) is crucial for metabolomics, but identifying unknown compounds is limited by spectral library coverage.
• Existing machine learning methods struggle with the complexity and sparsity of MS/MS spectral data and metabolite structures.

Purpose of the Study:

• To develop a novel machine learning method, ChemEmbed, for enhanced metabolite identification in MS/MS data.
• To address the limitations of current spectral annotation tools by integrating chemical structure information.

Main Methods:

• ChemEmbed utilizes multidimensional, continuous vector representations of chemical structures.
• MS/MS spectra are enhanced by merging data across multiple collision energies and incorporating calculated neutral losses.
• A convolutional neural network (CNN) processes the enhanced spectral and structural data.

Main Results:

• ChemEmbed correctly identifies the candidate metabolite in over 42% of cases and within the top five in over 76% of cases.
• The method demonstrated superior performance compared to SIRIUS 6 in CASMI 2016 and 2022 benchmarks.
• ChemEmbed successfully identified 25 previously unknown compounds in the Annotated Recurrent Unidentified Spectra (ARUS) dataset.

Conclusions:

• ChemEmbed offers a robust and scalable solution for accelerating metabolite identification in untargeted mass spectrometry.
• The integration of chemical structure embeddings with enhanced MS/MS spectra represents a significant advancement in computational metabolomics.