Cosine Similarity Algorithms Tested for Adulterant Detection in Drugs

In a recent study out of Central South University in Changsha, China, scientists tested different cosine similarity algorithms to test for illegal adulterants in drugs. Their findings were published in the Journal of Pharmaceutical and Biomedical Analysis (1).

Addictive substances, including alcohol, cigarettes and drugs | Image Credit: © monticellllo - stock.adobe.com

The dietary supplement and herbal medicine industry has grown rapidly in the past few decades. As a side effect of this growth, some manufacturers have added adulterants into their products to enhance their pharmacological effects, while severely risking consumers’ health in the process (1). Adulterants are substances, such as fillers, chemicals, and poisons, that are added to illicit drugs to enhance the potency and effects of a drug, all while bulking up the drug’s quantity and leading to increase profits. Common bulking agents can include starch, cellulose, sugars, and infectious organisms like bacteria and fungi. However, when added to street drugs, the dangers are magnified, including the risks of overdose and death. In one recent study of illicit drugs, 89–97% of the samples, which included fentanyl, cocaine, heroin, and more, contained adulterants (2). Side effects of adulterants in drugs can also include organ damage, infectious diseases, and cardiac arrest.

The risk associated with adulterants has led to a need for non-targeted screening methods for detecting adulteration in dietary supplements and herbal medicines reliably and rapidly. Normally, liquid chromatography–mass spectrometry (LC–MS) is used to detect adulteration, with high-resolution mass spectrometry (HRMS) gaining popularity in adulteration inspection compared to low-resolution mass spectrometry. HRMS has exceptional resolving power and can be used in different scanning modes that offer higher mass accuracy and better compound detection ability; additionally, HRMS can help identify unexpected or non-tracked compounds, which can aid the process.

Regarding non-tracked compounds, the most common method for mass spectral library matching is based on cosine similarity. While it can help in comparing nearly identical spectra, it is not as effective when query and reference compounds are structurally similar but have different specific chemical moieties, the researchers said. This has led to the creation of more advanced approaches, like Spec2Vec, MS2DeepScore, and entropy similarity. Spec2Vec is a novel spectral similarity score inspired by Word2Vec, a natural language processing algorithm (3). MS2DeepScore is a novel Siamese neural network to predict the structural similarity between two chemical structures solely based on their MS/MS fragmentation spectra (4). Entropy similarity involves matching tandem mass spectrometry (MS/MS) against experimental or in silico mass spectral libraries (5). In this experiment, Spec2Vec was used as the spectral similarity algorithm for identifying illegal adulterants, since it showed remarkable performance in detection capability and false discovery rates. The scientists’ goal was to optimize the performance of spectral similarity for a specific compound class by fine-tuning a pretrained Spec2Vec model. Additionally, the scientists used the chemical classification tool CANOPUS to address similarities in backbone structures between illegal adulterants and normal compounds found in herbal medicine, a phenomenon that can lead to false positives.

As a proof-of-concept, the scientists tested the system using glucocorticoids as potentially illicit adulterants. Following this experiment, the fine-tuned Spec2Vec model showed significant improvements in detection ability compared to the original model, achieving comparable performance levels to MS2DeepScore. Additionally, this fine-tuned Spec2Vec model showed notably less false positives compared to other similarity algorithms. This pipeline showed Spec2Vec as a highly effect and competitive means of detecting illegal adulterants, which can potentially enhance future analysis of large-scale MS data.

References

(1) Sheng, Y.; Xue, Y.; Wang, J.; Liu, S.; Jiang, Y. Nontargeted Screening Method for Detection of Illicit Adulterants in Dietary Supplements and Herbal Medicines Using UHPLC-QTOF-MS with Fine-Tuned Spec2Vec-Based Spectral Similarity and Chemical Classification Filter. J. Pharm. Biomed. Anal. 2024, 239, 115877. DOI: https://doi.org/10.1016/j.jpba.2023.115877

(2) Matta, N. Adulterants and Additives in Substances. American Addiction Centers 2023. https://americanaddictioncenters.org/ecstasy-abuse/adulterants-in-drugs-mdma (accessed 2024-3-15)

(3) Huber, F.; Ridder, L.; Verhoeven, S.; et al. Spec2Vec: Improved Mass Spectral Similarity Scoring Through Learning of Structural Relationships. PLOS Computational Biology 2021. https://doi.org/10.1371/journal.pcbi.1008724 (accessed 2024-3-15)

(4) Huber, F.; van der Burg, S.; van der Hooft, J. J. J.; Ridder, L. MS2DeepScore: A Novel Deep Learning Similarity Measure to Compare Tandem Mass Spectra. J. Cheminf. 2021, 13, 84. DOI: https://doi.org/10.1186/s13321-021-00558-4

(5) Li, Y.; Kind, T.; Folz, J.; Vaniya, A.; Mehta, S. S.; Fiehn, O. Spectral Entropy Outperforms MS/MS Dot Product Similarity for Small-Molecule Compound Identification. Nat. Methods 2021, 18, 1524–1531. DOI: https://doi.org/10.1038/s41592-021-01331-z