AIRI: Predicting Retention Indices and Their Uncertainties Using Artificial Intelligence.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Artificial Intelligence

Background:

• Kováts retention index (RI) is crucial for chemical identification in gas chromatography.
• Manual creation of RI libraries is time-consuming and labor-intensive.

Purpose of the Study:

• To develop a deep neural network for predicting RI values from chemical structures.
• To enhance chemical identification methods and spectral library quality using predicted RI values.
• To quantify prediction uncertainty for individual RI value predictions.

Main Methods:

• Utilized a deep neural network (Artificial Intelligence Retention Indices - AIRI network) to predict RI values from chemical structures.
• Employed an ensemble of 8 networks to estimate prediction uncertainty via standard deviation.
• Corrected predicted standard deviation to align with observed errors.

Main Results:

• The AIRI network achieved a mean absolute error of 15.1 and a 95th percentile absolute error of 46.5.
• Predicted RI values were integrated into NIST EI-MS spectral libraries.
• The uncertainty quantification method resulted in a standard deviation of 1.52 for Z scores.

Conclusions:

• Deep neural networks can accurately predict Kováts retention indices, significantly reducing the labor of library creation.
• The developed AI model enhances chemical identification accuracy and spectral library quality.
• Accurate uncertainty estimation is vital for practical application of AI-driven prediction models in chromatography.