General Intelligence-based Fragmentation (GIF): A framework for peak-labeled spectra simulation

Area of Science:

• Computational chemistry
• Metabolomics
• Artificial intelligence

Background:

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

Purpose of the Study:

• To introduce a novel framework, General Intelligence-based Fragmentation (GIF), for guiding LLMs in mass spectra simulation.
• To evaluate the performance of generalist LLMs in molecular fragmentation and intensity prediction using GIF.

Main Methods:

• Developed the GIF framework utilizing structured prompting, tagging, and iterative refinement.
• Fine-tuned pretrained LLMs and evaluated their performance on the MassSpecGym QA-sim dataset.
• Benchmarked GIF against other LLMs and deep learning models.

Main Results:

• GPT-4o and GPT-4o-mini achieved high cosine similarity (0.36 and 0.35) in spectral simulation.
• GIF outperformed other LLMs (GPT-5, Llama-3.1) and domain-specific models (ChemDFM).
• The framework demonstrated superior performance compared to several deep learning baselines.

Conclusions:

• GIF provides a structured approach for LLM-guided spectra simulation, enhancing molecular fragmentation analysis.
• LLMs, guided by GIF, can facilitate human-in-the-loop workflows and enable explainable reasoning in metabolomics.
• The study highlights the potential of systematic LLM guidance for complex scientific tasks in metabolomics.