Generating Molecular Fragmentation Graphs with Autoregressive Neural Networks.

Area of Science:

• Computational chemistry
• Metabolomics
• Cheminformatics

Background:

• Accurate tandem mass spectra prediction is crucial for metabolomic discovery and expanding reference libraries.
• Current computational methods (bond-breaking) are slow and inaccurate; neural networks are fast but lack interpretability and accuracy.

Purpose of the Study:

• To develop a physically grounded neural network approach for rapid and accurate tandem mass spectra prediction.
• To improve metabolite identification and enhance the interpretability of spectral prediction models.

Main Methods:

• A novel hybrid approach combining physical fragmentation principles with neural networks.
• Iterative simulation of molecular fragmentation events, scoring relevant molecular fragments.
• Evaluation using public and private standard spectral libraries.

Main Results:

• Achieved state-of-the-art prediction accuracy for tandem mass spectra.
• Demonstrated improved metabolite identification from candidate databases.
• Showcased higher interpretability compared to traditional and black-box neural network methods.
• Highlighted promise for natural product molecule elucidation.

Conclusions:

• The physically grounded neural approach offers a significant advancement in tandem mass spectra prediction.
• This method enhances the speed, accuracy, and interpretability of spectral prediction, aiding metabolomic research.
• The approach shows particular potential for analyzing complex natural product structures.