Evaluating the generalizability of graph neural networks for predicting collision cross section

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Ion Mobility coupled with Mass Spectrometry (IM-MS) provides molecular size and shape information via collision cross-section (CCS) values.
• Accurate in silico prediction of CCS is crucial due to limited experimental data, driving the development of machine learning (ML) models.

Purpose of the Study:

• To evaluate the performance of state-of-the-art Graph Neural Networks (GNNs) for predicting CCS values.
• To investigate the generalization capabilities of GNNs on structurally novel chemical spaces.
• To introduce methods for improving the reliability and generalizability of in silico CCS predictions.

Main Methods:

• Benchmarking state-of-the-art GNNs using the largest publicly available CCS dataset.
• Evaluating model performance on both familiar and structurally novel chemical regions.
• Developing and testing the Mol2CCS approach, incorporating molecular fingerprints, descriptors, and molecule types.
• Implementing confidence models to enhance CCS estimate reliability.

Main Results:

• GNNs demonstrate high accuracy for CCS prediction within chemical spaces similar to their training data.
• Model performance significantly degrades when applied to structurally novel molecules.
• The Mol2CCS approach shows partial improvement in generalization by including additional molecular features.
• Confidence models can enhance the reliability of in silico CCS predictions.

Conclusions:

• Current ML models for CCS prediction are reliable primarily within their training data's chemical space.
• Significant challenges remain in achieving robust generalization for predicting CCS of novel molecular structures.
• Further development of ML models and the release of more diverse CCS datasets are essential for advancing in silico structural analysis.