Rapid Approximate Subset-Based Spectra Prediction for Electron Ionization-Mass Spectrometry.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Mass spectrometry is crucial for identifying compounds and determining chemical structures.
• Accurate mass spectral data is essential for compound identification.
• Computational prediction of mass spectra can expand spectral databases for unmeasured molecules.

Purpose of the Study:

• To present a novel computational method for predicting electron ionization-mass spectra (EI-MS) of small molecules.
• To improve the accuracy and utility of spectral databases for chemical analysis.
• To develop deep learning models that can predict mass spectra from chemical structures.

Main Methods:

• Developed two deep learning models: FormulaNet and SubsetNet, combined as rapid approximate subset-based spectra prediction (RASSP).
• FormulaNet predicts subformulae, while SubsetNet predicts vertex subsets of molecular graphs.
• Models were trained and evaluated using the NIST 2017 Mass Spectral Library and PubChem data.

Main Results:

• FormulaNet achieved 92.9% weighted dot product accuracy and 98.0% recall (top 10).
• SubsetNet demonstrated strong generalization, particularly in high-resolution, low-data scenarios.
• The best model reduced spectral database lookup error rate by 2.9x (from 5.7% to 2.0% in top 10).

Conclusions:

• RASSP significantly enhances the accuracy of EI-MS prediction compared to previous methods.
• The developed models offer superior performance and generalization for spectral database applications.
• Freely accessible source code and 73.2 million predicted spectra will benefit the scientific community.