MSBERT: Embedding Tandem Mass Spectra into Chemically Rational Space by Mask Learning and Contrastive Learning

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Bioinformatics

Background:

• Tandem mass spectrometry (MS/MS) generates vast spectral data crucial for compound identification.
• Liquid chromatography-mass spectrometry (LC-MS) experiments produce large datasets requiring efficient analysis.
• Current methods for MS/MS spectral interpretation face challenges in speed and accuracy.

Purpose of the Study:

• To develop an efficient method for compound identification using MS/MS spectral data.
• To create effective embeddings for MS/MS spectra through self-supervised learning.
• To enhance capabilities in spectral library matching and analogous compound searching.

Main Methods:

• Proposed MSBERT, a model based on self-supervised learning and transformer encoder architecture.
• Utilized mask learning and contrastive learning strategies on MS/MS spectra.
• Trained and tested MSBERT on diverse datasets including GNPS, MoNA, and MTBLS1572.

Main Results:

• MSBERT demonstrated superior performance in library matching and analogous compound searching.
• Achieved higher recall rates (0.7871, 0.8950, 0.9080 at 1, 5, 10) compared to Spec2Vec and DreaMS.
• Validated embedding rationality through t-SNE visualization, structural similarity, and clustering analyses.

Conclusions:

• MSBERT offers enhanced compound identification capabilities from MS/MS data.
• The developed embeddings are rational and improve spectral analysis.
• A user-friendly web server and open-source code are provided for accessibility.