The use of ion mobility separation (IMS) in conjunction with high-resolution mass spectrometry has proved to be a reliable and useful technique for the characterization of small molecules from plastic products. Collision cross-section (CCS) values derived from IMS can be used as a structural descriptor to aid compound identification. One limitation of the application of IMS to the identification of chemicals from plastics is the lack of published empirical CCS values. As such, machine learning techniques can provide an alternative approach by generating predicted CCS values. Herein, experimental CCS values for over a thousand chemicals associated with plastics were collected from the literature and used to develop an accurate CCS prediction model for extractables and leachables from plastic products. The effect of different molecular descriptors and machine learning algorithms on the model performance were assessed. A support vector machine (SVM) model, based on Chemistry Development Kit (CDK) descriptors, provided the most accurate prediction with 93.3% of CCS values for [M + H]+ adducts and 95.0% of CCS values for [M + Na]+ adducts in testing sets predicted with <5% error. Median relative errors for the CCS values of the [M + H]+ and [M + Na]+ adducts were 1.42 and 1.76%, respectively. Subsequently, CCS values for the compounds in the Chemicals associated with Plastic Packaging Database and the Food Contact Chemicals Database were predicted using the SVM model developed herein. These values were integrated in our structural elucidation workflow and applied to the identification of plastic-related chemicals in river water. False positives were reduced, and the identification confidence level was improved by the incorporation of predicted CCS values in the suspect screening workflow.
Keywords: collision cross-section; extractables; ion mobility; leachables; machine learning; plastic products.