Can Large Language Models Translate MS/MS into Molecular Caption?

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Bioinformatics

Background:

• Mass spectrometry (MS) is crucial for molecular discovery but faces challenges in interpreting tandem mass spectra (MS/MS) due to database reliance and limited structural detail.
• Current methods for MS/MS interpretation often require extensive databases and struggle with resolving complex molecular structures.

Purpose of the Study:

• To investigate the potential of large language models (LLMs) to directly translate MS/MS spectra into chemically meaningful molecular descriptions.
• To develop and evaluate a novel framework, MS2LLM, for this spectral-to-description translation task.

Main Methods:

• Developed MS2LLM, a framework representing spectra and molecular structures as natural language.
• Employed instruction tuning to train the model on the correspondence between spectral data and molecular descriptions.
• Generated hierarchical molecular descriptions (functional groups, substructures, chemical classes) directly from MS/MS spectral inputs.

Main Results:

• MS2LLM demonstrated superior performance compared to general-purpose LLMs and conventional spectral learning methods in descriptive accuracy and chemical classification across multiple datasets.
• The model successfully generated interpretable outputs, capturing structural semantics rather than exact structures.
• Achieved enhanced structural inference capabilities for unknown molecules.

Conclusions:

• LLMs, specifically through the MS2LLM framework, offer a promising new avenue for interpreting MS/MS spectra.
• MS2LLM provides a complementary paradigm for structural inference, enhancing the analysis of unknown molecules in molecular discovery.
• The approach facilitates a deeper understanding of spectral data by generating chemically relevant molecular descriptions.