General Intelligence-Based Fragmentation (GIF): A Framework for Peak-Labeled Spectra Simulation.

Area of Science:

• Metabolomics
• Computational Chemistry
• Artificial Intelligence

Background:

• Metabolomics research is limited by low spectral annotation rates, hindering progress despite advanced tools.
• Large language models (LLMs) show promise for scientific applications, including complex tasks like mass spectra annotation.

Purpose of the Study:

• To introduce General Intelligence-based Fragmentation (GIF), a framework for guiding pretrained LLMs in mass spectra simulation.
• To evaluate LLMs' reasoning capabilities in molecular fragmentation and intensity prediction using structured prompting.

Main Methods:

• Developed GIF, a framework employing tagging, structured I/O, system prompts, and iterative refinement for LLM guidance.
• Fine-tuned and evaluated generalist LLMs on the MassSpecGym QA-sim dataset for spectra simulation.
• Benchmarked GIF against other LLMs (GPT-5, Llama-3.1) and domain-specific models (ChemDFM).

Main Results:

• GPT-4o and GPT-4o-mini achieved high cosine similarity (0.36 and 0.35) in simulated vs. true spectra.
• GIF outperformed several deep learning baselines and other leading LLMs in spectra simulation accuracy.
• The framework demonstrated superior performance compared to GPT-5 and ChemDFM.

Conclusions:

• GIF provides a structured approach to LLM prompting for scientific tasks like molecular fragmentation.
• LLMs, guided by GIF, show significant potential for improving spectra simulation and enabling explainable AI in metabolomics.
• The framework supports human-in-the-loop workflows, enhancing the utility of LLMs in scientific discovery.