In-silico Tandem Mass Spectral Libraries for the Blood Exposome and Metabolomics Workbench Databases: Application of Gemini Large Language Model

In the human body, various endogenous metabolites, nutrients, and xenobiotic chemicals exist. These substances can be identified within a biospecimen by employing high-resolution mass spectrometry instruments, which generates metabolomic and exposomic datasets. A crucial element in these processes is the utilization of mass spectral libraries to assign chemical information to experimental data.

In this data descriptor, researchers present mass spectral libraries for human-relevant small molecules, which were produced utilizing the Google Gemini large-language model. The team employed an optimized chain-of-thought prompt and instructions to generate mass spectra for each compound at high, medium, and low energy fragmentation levels.

Each spectrum was reported in the NIST MSP text format to facilitate easy integration with metabolomics software. The libraries encompass 91,231 unique 2D structures derived from the Blood Exposome Database and the Metabolomics Workbench data repository. Technical validation was conducted by comparing the Gemini-generated spectra against the NIST 23 mass spectral database, yielding an accuracy rate of 55%.

The use of Large Language Models (LLMs) such as Gemini demonstrates their potential as an accessible and useful tool for expanding in-silico spectral libraries. This expansion aids in the annotation of compounds within the High-Resolution Mass Spectrometry (HRMS) data of human biospecimens, thereby potentially expediting the process of compound annotation in metabolomics and exposomics.