Predicting Tandem Mass Spectra of Small Molecules Using Graph Embedding of Precursor-Product Ion Pair Graph

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Bioinformatics

Background:

• Metabolomics relies on high-quality tandem mass spectrometry (MS/MS) data for accurate compound identification.
• Current MS/MS prediction methods face challenges in accuracy, resolution, and structural correlation.
• Developing robust MS/MS prediction is crucial for advancing metabolomics research.

Purpose of the Study:

• To develop a novel MS/MS prediction method, PPGB-MS2, for enhanced metabolomics identification.
• To improve the accuracy, resolution, and structural correlation of MS/MS fragmentation prediction.
• To leverage graph neural networks (GNNs) for precise fragment intensity prediction.

Main Methods:

• Introduced precursor-product ion pair graph bags (PPGBs) for uniform representation of molecular structures and fragmentation data.
• Transformed MS/MS prediction into fragment intensity prediction using PPGBs.
• Utilized graph neural networks (GNNs) for machine learning-based MS/MS fragment intensity prediction.
• Trained and evaluated the model using [M+H]+ and [M-H]- data from the NIST 20 tandem MS database.

Main Results:

• Achieved an average cosine similarity of 0.71 on the test set, outperforming classical MS/MS prediction methods.
• Demonstrated high-resolution MS/MS prediction capabilities through effective fragment-structure correspondence.
• Successfully integrated chemical structure information into machine learning models for prediction.

Conclusions:

• PPGB-MS2 offers a significant advancement in MS/MS prediction accuracy and reliability for metabolomics.
• The PPGB representation and GNNs provide a powerful framework for predicting MS/MS fragmentation patterns.
• This method enhances the ability to identify metabolites by improving the quality of spectral data.