Predicting differential ion mobility behaviour in silico using machine learning.

Area of Science:

• Analytical Chemistry
• Physical Chemistry

Background:

• Differential mobility spectrometry (DMS) is increasingly used in analytical workflows.
• Current DMS method development relies on time-consuming manual optimization of separation conditions.
• Accurate prediction of differential ion mobility is needed to streamline this process.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting ion mobility and dispersion curves in DMS.
• To reduce the manual effort and time required for DMS method development.

Main Methods:

• A random-forest based ML model was trained using DMS data from 409 cationic analytes.
• The model utilized m/z and ion-neutral collision cross section (CCS) as input features.
• Model performance was evaluated by comparing predicted dispersion curves to experimental data, with and without additional low-voltage data points.

Main Results:

• The initial ML model predicted dispersion curves with a mean absolute error (MAE) of ≤ 2.4 V.
• Incorporating two additional data points at lower separation voltages (SVs) improved accuracy to MAE ≤ 1.2 V.
• The guided ML routine effectively captured Type A and B ion behaviors, significantly improving predictions for common Type C ions.

Conclusions:

• Machine learning offers a powerful approach to predict differential ion mobility and optimize DMS parameters.
• The developed ML model significantly reduces the time and effort associated with DMS method development.
• Further refinement of the ML model with targeted data can enhance prediction accuracy for diverse ion behaviors.