Effect-directed analysis of genotoxicants in food packaging based on HPTLC fractionation, bioassays, and toxicity prediction with machine learning.

Many chemicals in food packaging can leach as complex mixtures to food, potentially including substances hazardous to consumer health. Detecting and identifying all of the leachable chemicals are impractical with current analytical instrumentation and data processing methods. Therefore, our work aims to expand the analytical toolset for prioritizing and identifying chemical hazards in food packaging. We used a high-performance thin-layer chromatography (HPTLC)-based bioassay to detect genotoxic fractions in paperboard packaging. These fractions were then processed with non-targeted liquid chromatography high-resolution mass spectrometry (LC-HRMS/MS) and machine learning-based toxicity prediction (MLinvitroTox). The HPTLC bioassay detected four genotoxic zones in extracts of the paperboard. One-dimensional HPTLC separation and targeted fraction collection reduced the number of chemical features extracted from paperboard and detected with LC-HRMS by at least 98% (from 1695-2693 to 14-50). The entire process was successful for spiked genotoxic chemicals, which were correctly prioritized in the fractionation and non-target analysis workflow. The native chemical with the strongest genotoxicity signal was identified with a suspect list as 5-chloro-2-methyl-4-isothiazolin-3-one and confirmed with LC-HRMS/MS and HPTLC bioassay. Toward identification of the remaining unknown genotoxicants, two-dimensional HPTLC further reduced the number of chemical features. Genotoxicity predictions with MLinvitroTox based on molecular fingerprints of the unknown signals derived from their MS2 fragmentation spectra helped prioritize two chemical features and suggested candidate structures. This work demonstrates strategies for using HPTLC, HRMS, and toxicity prediction to help identify toxicants in food packaging.