Collision Cross Section Prediction with Molecular Fingerprint Using Machine Learning.

Area of Science:

• Analytical Chemistry
• Environmental Science
• Computational Chemistry

Background:

• High-resolution mass spectrometry (HRMS) coupled with ion mobility spectrometry (IMS) is crucial for non-target screening (NTS) of emerging contaminants.
• Current identification methods in NTS rely on spectral libraries and databases, often lacking orthogonal data for confident small molecule identification.
• Collision cross section (CCS) values provide an additional dimension for identification, but prediction methods are limited in scope and cross-platform validation.

Purpose of the Study:

• To develop and validate robust machine learning models for predicting collision cross section (CCS) values.
• To enhance the accuracy and reliability of small molecule identification in non-target screening (NTS) applications.
• To provide a tool applicable across different ion mobility spectrometry (IMS) platforms for improved contaminant monitoring.

Main Methods:

• Development of two random forest machine learning models for CCS prediction.
• Training and testing models using over 13,324 CCS values from diverse sources, including six laboratories and PubChem.
• Utilizing chemical super classes and molecular fingerprints as input features for the prediction models.

Main Results:

• Achieved test accuracy exceeding 0.85 for all developed prediction models.
• Reported a median relative residual of approximately 2.2% for CCS predictions.
• Demonstrated the models' capability to be applied across various IMS platforms.

Conclusions:

• The developed machine learning models significantly improve the accuracy of collision cross section (CCS) prediction.
• These models offer a valuable tool for enhancing small molecule identification confidence in non-target screening (NTS).
• The cross-platform applicability of the models aids in reducing false positives and advancing contaminant monitoring.