Fast screening and identification of illegal adulteration in dietary supplements and herbal medicines using molecular networking with deep-learning-based similarity algorithms.

Liquid chromatography-high-resolution mass spectrometry (LC-HRMS) is a powerful analytical tool used for adulteration inspection. Nevertheless, it is a challenging task to identify illegal adulterants that are not included in the library or are unexpected from large MS data. Molecular networking is a good tool for exploring, visualizing, and organizing MS/MS spectra, and moreover, it employs shifted peak match to calculate spectral similarity, making it capable of identifying adulteration that is not included in the library. The key of molecular networking is spectral similarity algorithms, and therefore, in this study, we compared the performance of four cutting-edge similarity algorithms, modified cosine similarity (shifted peak match), entropy similarity, and two deep-learning-based algorithms, MS2DeepScore and Spec2Vec, in building molecular networking for identification of adulteration that is not included in the library. We conducted an analysis of excluded-query-compound on all MS/MS spectra in test library and performed a large-scale false discovery rate estimation to investigate whether the spectral similarity calculated by each algorithm could represent the actual structural similarity well. The obtained results demonstrated Spec2Vec exhibited good performance in both detection capability and false discovery rate. Further comprehensive evaluation of the performance of Spec2Vec in the identification of adulteration that is not included in the library or is unexpected in different matrices and in application to real samples proved the approach studied here is a promising and powerful tool for adulterant inspection and improved the capability of analyzing large MS data.