Supervised Machine Learning and Graph Neural Networks to Predict Collision Cross-Section Values of Aquatic Dissolved...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Analytical Chemistry

Background:

• Collision Cross-Section (CCS) prediction is crucial for identifying molecules in complex environmental mixtures.
• Existing methods for molecular identification can be limited in accuracy and scope.

Purpose of the Study:

• To develop and evaluate machine learning and deep learning models for predicting CCS values of diverse dissolved organic molecules.
• To enhance molecular classification and improve contaminant detection in environmental samples.

Main Methods:

• Evaluated eight supervised regression models (e.g., Random Forest, SVR, Voting Regressor) and a Graph Neural Network (GNN).
• Models were trained using molecular fingerprints (SMILES) and structural descriptors (m/z, O/C, H/C, AImod, DBE).
• Validated model performance using High-Resolution Mass Spectrometry (HRMS) data from the Arctic Ocean.

Main Results:

• Model performance varied by molecular class, with specific models excelling for different compound types (e.g., Voting Regressor for carbohydrates, Random Forest for hydrocarbons).
• The Graph Neural Network (GNN) demonstrated consistent, competitive accuracy across all molecular classes.
• Validation confirmed the models' predictive power for accurate molecular structure selection.

Conclusions:

• A robust, data-driven framework for CCS prediction was established, enhancing molecular classification.
• The developed models significantly improve the precision of identifying molecular structures and detecting contaminants in environmental analyses.