The investigators produced in-silico tandem mass spectral libraries for human-relevant small molecules using the Gemini large language model. The libraries encompass 91,231 unique 2D structures sourced from two repositories: the Blood Exposome Database and the Metabolomics Workbench. The preprint reports that each compound received predicted spectra, making this a large-scale computational effort focused on chemicals that appear in human biospecimens.
Coverage at this scale aims to address a common bottleneck in untargeted high-resolution mass spectrometry workflows: the limited availability of experimental tandem mass spectra for many biologically relevant molecules. The authors did not provide per-compound breakdowns in the preprint beyond the total structure count.
Each generated spectrum was exported in the NIST MSP text format, which the study team selected to ease integration with existing metabolomics software. Spectra were produced at three fragmentation energy levels: high, medium and low. The preprint specifies those three energy tiers but does not include instrument-specific parameters or the exact definitions of the energy windows used for generation. Data in MSP format should permit users to import the libraries into many common spectral-matching pipelines.
The authors applied an optimized chain-of-thought prompt and instruction set to steer the Gemini model toward producing tandem mass spectra. The preprint reports that this prompt optimization was a component of the workflow, but it did not publish the full prompt text or the precise computational pipeline in the main abstract. The study indicates an LLM-centric approach rather than physics-based fragmentation simulation, and that choice shaped how spectra were synthesized at the three energy levels.
The preprint links to external data and code resources (a Zenodo record is cited), which may provide further implementation details. The manuscript does not, however, present detailed runtime metrics, compute resources used, or step-by-step reproducible commands in the abstract.
Technical validation compared the Gemini-generated spectra against the NIST 23 mass spectral database. The authors reported a 55% accuracy rate from that comparison. The preprint does not define the exact matching threshold, scoring algorithm, or validation subset used to derive the 55% figure within the abstract text; those methodological specifics may appear in the full manuscript or linked materials.
A 55% concordance with an established, experimentally derived spectral database signals partial agreement between model-generated and reference spectra. The preprint frames this result as evidence that large language models can be a useful and accessible tool to expand in-silico spectral resources for HRMS compound annotation, while implicitly acknowledging that performance is not equivalent to experimental spectra.
The authors included a Zenodo data/ code record and reported funding from NIH/NIEHS (several grant numbers listed) and NIH/NCATS. The preprint provides institutional affiliations for the authors at Icahn School of Medicine at Mount Sinai and supplies a corresponding author email. The manuscript notes no competing interests.
Users seeking to deploy these libraries should consult the linked Zenodo record and the full preprint PDF for downloadable MSP files and any accompanying metadata files that the authors made available.
This work appears as a bioRxiv preprint and has not been peer reviewed. The preprint explicitly states that the article has not been certified by peer review. Reported limitations in the abstract include the validation outcome (55% accuracy against NIST 23), which indicates that model-generated spectra still diverge from experimental references in many cases. The source did not report detailed prompt text or full validation methodology in the abstract, and it did not include instrument-specific fragmentation definitions for the high/medium/low energy tiers.
Remaining open questions the study raises, based on the preprint text, include: how scoring and match thresholds were chosen for validation; whether per-class or per-chemical-family accuracy differs from the aggregate 55% figure; and how these in-silico spectra perform in real-world metabolomics pipelines for identification rates and false discovery control. Those items are not answered in the abstract and would require consultation of the full preprint and associated data/code links. The article closes on practical transparency: data and code pointers are provided, and the preprint status and validation results frame the libraries as a resource with measurable but incomplete concordance to experimental spectra.