Predicting Collision Cross-Section Values for Small Molecules through Chemical Class-Based Multimodal Graph Attention...

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Biochemistry

Background:

• Collision cross-section (CCS) libraries are crucial for compound identification in metabolomics.
• Accurate CCS prediction is challenging due to the vast structural diversity of small molecules.
• Existing computational methods struggle with precise CCS value prediction.

Purpose of the Study:

• To develop an advanced machine learning model for accurate CCS value prediction.
• To improve compound identification in metabolomics using predicted CCS values.
• To create a user-friendly tool for accessing the predictive model.

Main Methods:

• Developed a machine learning model (MGAT-CCS) integrating graph attention networks and multimodal molecular representations.
• Trained and validated the model on diverse chemical classes, including lipids and metabolites.
• Applied the model to real-world metabolomics data from various biological samples.

Main Results:

• MGAT-CCS demonstrated superior performance compared to other ML models for CCS prediction.
• Achieved low median relative errors: 0.47%/1.14% (positive/negative mode) for lipids and 1.40%/1.63% for metabolites.
• Reduced false metabolite candidates by approximately 25% in real-world metabolomics data.
• Developed a publicly available web server for MGAT-CCS.

Conclusions:

• MGAT-CCS significantly advances the accuracy of CCS value prediction.
• The model facilitates more reliable small molecule identification in metabolomics.
• The accessible web server promotes wider adoption and application in the scientific community.