IMPACT-4CCS: Integrated Modeling and Prediction Using Ab Initio and Trained Potentials for Collision Cross Sections

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Analytical Chemistry

Background:

• Collision cross section (CCS) is crucial for identifying and classifying per- and polyfluoroalkyl substances (PFAS).
• Existing computational methods face challenges with large molecules, increasing PFAS candidates, and generalizing machine learning models to diverse structures.
• Accurate and timely CCS prediction is essential for environmental monitoring and risk assessment of PFAS.

Purpose of the Study:

• To develop a novel computational workflow ensemble, IMPACT4-CCS, for accelerating accurate CCS prediction of PFAS molecules.
• To improve the accuracy and generalizability of CCS prediction, particularly for emerging PFAS subclasses and flexible molecular structures.
• To demonstrate the advantage of integrating ab initio methods with machine learning over purely data-driven approaches for CCS analysis.

Main Methods:

• Developed IMPACT4-CCS, a computational workflow combining ab initio calculations with machine learning.
• Validated the method's accuracy against a test set of 100 molecules.
• Assessed performance on specific PFAS subclasses, including nH-perfluoroalkyl carboxylic acids (nH-PFCA), and evaluated the capability to capture structural dynamics like hydrogen bridging.

Main Results:

• IMPACT4-CCS achieves accuracy comparable to current machine learning approaches.
• The method shows superior accuracy for emerging PFAS subclasses, such as nH-PFCA, outperforming other methods that overestimate CCS values.
• IMPACT4-CCS is the first method demonstrated to capture structural dynamics (hydrogen bridging) in large, flexible PFAS molecules.

Conclusions:

• IMPACT4-CCS offers a significant advancement in the accurate and efficient prediction of CCS for PFAS.
• Integrating machine learning with traditional computational methods provides greater accuracy and generalizability than relying solely on machine learning on molecular graphs.
• The developed workflow has potential applications in expanding nontarget analysis for vast PFAS datasets, such as the OECD PFAS list.