Abstract:

Research on secondary metabolites is of great significance for plant growth and development, environmental adaptation, as well as human health and drug development. Liquid chromatography-mass spectrometry（LC-MS）has become the preferred strategy for secondary metabolism research. However, the annotation of metabolite structures is still hindered by the insufficient coverage of standard spectral libraries. Given that the coverage of metabolite structure databases far exceeds that of standard spectral libraries, establishing the association between metabolite structures and mass spectra through artificial intelligence methods to search molecular structure databases based on mass spectrometry data is an effective approach to address this issue. This paper reviews three strategies for establishing the association between metabolite structures and mass spectra using deep learning techniques and bioinformatics methods, including structure-to-spectrum, spectrum-to-structure, and known-to-unknown strategies. It also introduces the rationale and representative methods for each strategy. For each strategy, the paper discusses the advantages and limitations of its algorithms, as well as the challenges that may be encountered in practical applications. Additionally, the paper explores that factors should be considered when developing new algorithms and conducting benchmark tests, and how these factors may affect the evaluation of algorithms. Finally, the paper points out that integrating more orthogonal information is a future direction for achieving more accurate metabolite annotation.

Key words: secondary metabolism, bioinformatics, mass spectrometry analysis, artificial intelligence