Sodium adduct formation with graph-based machine learning can aid structural elucidation in non-targeted LC/ESI/HRMS

Area of Science:

• Analytical Chemistry
• Computational Chemistry

Background:

• Non-targeted screening using LC/ESI/HRMS is crucial for identifying unknown compounds.
• Distinguishing between isomeric compounds remains a significant challenge in structural elucidation.
• Differences in ionic species formed during electrospray ionization (ESI) offer a potential avenue for isomer differentiation.

Purpose of the Study:

• To investigate adduct formation in positive ESI mode for improved isomer identification.
• To develop and validate a predictive model for sodium adduct formation in ESI/HRMS.
• To assess the utility of predicted sodium adduct formation for distinguishing isomers in large databases.

Main Methods:

• Studied adduct formation for 94 small molecules using ion mobility spectrometry and collision cross-section analysis.
• Developed a support vector machine classifier with polynomial kernels trained on diverse datasets.
• Utilized graph-based functional group connectivity and PubChem fingerprints as model features.

Main Results:

• Achieved 74.7% accuracy (70.0% balanced accuracy) in predicting sodium adduct formation on an independent dataset.
• The classification algorithm was applied to the SusDat database.
• Sodium adduct formation probability provided additional selectivity for approximately 25% of exact masses.

Conclusions:

• The developed machine learning model accurately predicts sodium adduct formation in ESI/HRMS.
• Predicted sodium adduct formation offers practical utility for enhancing selectivity in non-targeted screening.
• This approach aids in the structural assignment of isomeric compounds, overcoming a key limitation in current methods.