Highly accurate and large-scale collision cross sections prediction with graph neural networks

Area of Science:

• Computational chemistry
• Analytical chemistry
• Machine learning

Background:

• Collision cross section (CCS) values from ion mobility spectrometry aid compound identification.
• Accurate CCS prediction is crucial for improving analytical method reliability.

Purpose of the Study:

• To develop an accurate and reliable method for predicting CCS values directly from molecular structures.
• To create a large-scale in-silico database of CCS values for diverse compounds.

Main Methods:

• Developed SigmaCCS, a graph neural network model utilizing 3D conformers.
• Trained and validated the model on over 5,000 experimental CCS values.
• Employed model-agnostic interpretation and visualization for chemical rationality assessment.

Main Results:

• SigmaCCS achieved a high coefficient of determination (0.9945) and low median relative error (1.1751%) on the test set.
• Generated an in-silico database of 282 million CCS values for 94 million compounds across three adduct types.
• The prediction method demonstrated chemical rationality through interpretation and visualization.

Conclusions:

• SigmaCCS is an accurate, rational, and readily available method for predicting CCS values from molecular structures.
• The developed tool and database significantly advance compound identification capabilities in ion mobility spectrometry.
• Publicly available source code facilitates broader adoption and further research.