Machine Learning-Assisted Key Toxicant Identification from Mass Spectra: A Case Study of AhR Agonists in Polypropylene Face Masks

Abstract

Abstract Identifying key toxicants that drive biological effects in complex samples remains a major challenge for environmental analytical chemistry. Here, we developed an AhR-oriented machine-learning model, EIMSFrag-AhR, to predict AhR agonistic activity directly from GC-EI-MS spectral data and combined it with the previously reported LC-MS/MS-based MSFrag-Tox model to support toxicant identification in polypropylene face masks. The two models were applied to GC-EI-HRMS and LC-ESI-HRMS data sets from six highly active mask samples, prioritizing 46 potential AhR-active features. Subsequent structural annotation, authentic standard confirmation, and bioassay validation identified 12 AhR agonists. Six chemicals, including the frequently detected intermediate 2,4-dimethylaniline (GC data set), the UV absorber 2,4-dihydroxybenzophenone (GC data set), and the solely detected corrosion inhibitor 4-methyl-1H-benzotriazole (LC data set), were identified here as AhR agonists for the first time. The identified toxicants were broadly consistent with the key structural features highlighted by the two models. Detection frequencies ranged from 1 to 41 mask samples, suggesting both sample-specific and more widespread sources. Contribution analysis based on fitted AC30-derived TEFs showed that the validated compounds contributed less than 0.1% of the total AhR effect in one highly active mask. This proof-of-concept study demonstrates that machine learning-assisted spectrum-to-bioactivity prediction can support toxicant prioritization and identification in complex real samples.