Predicting Kováts Retention Indices Using Graph Neural Networks.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Kováts retention index is a crucial, near-universal descriptor for gas chromatography (GC) retention times.
• Experimental determination of Kováts retention indices is extensive, with curated datasets like the NIST 20 library being invaluable.
• Predictive models for retention indices can significantly aid chemical analysis and compound identification.

Purpose of the Study:

• To train and evaluate a graph neural network (GNN) model for predicting Kováts retention indices.
• To compare the GNN model's performance against existing methods, including convolutional neural networks and group contribution approaches.
• To demonstrate the efficacy of deep learning for chemical data analysis using the NIST 20 library.

Main Methods:

• Utilized the NIST 20: GC Method/Retention Index Library as the primary data source.
• Developed and trained a graph neural network model on curated experimental retention index data.
• Compared the predictive accuracy of the GNN model against established methods using mean unsigned error (MUE).

Main Results:

• The GNN model achieved a mean unsigned error of 28 index units for Kováts retention index prediction.
• This represents a significant improvement over previous best results, such as a convolutional neural network with an MUE of 44.
• The GNN model outperformed a group contribution approach (MUE of 114) using the same input data.

Conclusions:

• Deep learning methodologies, specifically graph neural networks, offer powerful predictive capabilities for chemical data.
• The developed GNN model provides a highly accurate and data-driven approach for predicting Kováts retention indices from the NIST 20 library.
• Systematic, data-driven approaches significantly outperform traditional methods for retention index prediction in gas chromatography.