Predicting the Predicted: A Comparison of Machine Learning-Based Collision Cross-Section Prediction Models for Small...

Area of Science:

• Computational chemistry
• Analytical chemistry
• Metabolomics

Background:

• Machine learning (ML) is increasingly applied in -omics research due to large datasets.
• ML models predict tandem mass spectrometry and retention time in metabolomics.
• New ML models predict collision cross-section (CCS) using ion mobility data, but face limitations.

Purpose of the Study:

• To compare existing ML-based CCS prediction models using the METLIN-CCS dataset.
• To evaluate ML models trained with fingerprints and simple linear models.
• To analyze the impact of training data structural diversity on CCS prediction accuracy and metabolite annotation.

Main Methods:

• Comparison of four existing ML-based CCS prediction models.
• Evaluation against linear models and ML models using fingerprints.
• Analysis of prediction accuracy using the METLIN-CCS dataset.
• Exploration of training data structural diversity's role.

Main Results:

• Existing ML models demonstrated limited accuracy for routine metabolomics analysis.
• Prediction accuracy improved significantly only when models were trained on large sets of structurally similar molecules.
• Current models have restricted capabilities for metabolite annotation using CCS values.

Conclusions:

• Current ML-based CCS prediction models are insufficient for routine metabolomics.
• Training data heterogeneity, with large hubs of similar molecules, is crucial for enhanced prediction and annotation.
• Future advancements require diverse datasets focusing on structurally related compound classes.