Structural elucidation using GCxGC-TOFMS and machine learning for unknown metabolites in HeLa cell

Importance of the Topic

Metabolomics provides a comprehensive view of small molecules produced by living systems during biological activities and enables deeper insights into cellular processes.
Advanced gas chromatography coupled with high‐resolution time‐of‐flight mass spectrometry (GC‐HRTOFMS) combined with machine learning addresses critical challenges in identifying unknown metabolites not present in existing spectral libraries.

Aims and Overview of the Study

This work demonstrates a workflow integrating GCxGC‐TOFMS (JMS‐T2000GC “AccuTOF GC‐Alpha”) with AI‐based software (msFineAnalysis AI).
The objectives are to generate a predicted EI spectral database of over 200 million entries, improve structural elucidation accuracy for unknowns, and apply the method to HeLa cell metabolome profiling.

Methods and Instrumentation

Sample Analysis Workflow:

• Two‐dimensional GC separation using a BPX5 first column (30 m×0.25 mm, 0.25 µm) and Rxi‐17Sil MS second column (3.4 m×0.15 mm, 0.15 µm) with thermal modulation.
• Time‐of‐flight MS detection with EI, CI, PI, FI ionization modes (resolving power >30 000, mass accuracy ±1 ppm).
• Soft ionization (FI) to capture molecular ions and enable formula determination via isotope patterns.
• Data processing through msFineAnalysis AI:
  
  • Initial NIST EI spectral database search (347 100 measured compounds).
  • Deconvolution to extract EI fragment ions, derive neutral losses, and filter by formula from soft ionization data.
  • AI‐predicted EI and retention index (RI) matching against a library of 200 million compounds.
  • Substructure prediction and ranking of candidate structures by cosine similarity.

Key Results and Discussion

AI Model Performance:

• Average cosine similarity of 0.86 between predicted and measured spectra for 10 000 held‐out compounds.
• Evolution of prediction accuracy from 0.72 (Ver.1, 2022) to 0.86 (Ver.3, 2025).
• Structural prediction success rates improved from 22% to 56% for top‐hit identification, with 82% and 87% correct candidates in top 5 and top 10 lists, respectively.

HeLa Cell Application:

• Detection of over 800 metabolites by GCxGC‐TOFMS, including amino acids, sugars, organic acids, and phosphorylated nucleotides.
• Successful identification of N-methyl-UMP (a metabolite absent from standard NIST libraries) through integrated EI/soft‐ionization analysis and AI‐driven structure ranking.

Benefits and Practical Applications

The combined GC‐TOFMS and AI workflow:

• Enables robust, non‐targeted metabolite profiling with high reproducibility and structural confidence.
• Circumvents limitations of existing spectral libraries by generating in-silico predicted spectra.
• Facilitates rapid molecular formula determination and structural elucidation for unknown compounds.

Future Trends and Opportunities

Expanding AI‐predicted databases to include additional chemical classes (e.g., polymers, natural products).
Integrating multi‐omics data (proteomics, transcriptomics) with metabolomics for holistic system analyses.
Advancing AI algorithms for improved retention index prediction and fragment interpretation.
Developing real-time workflows for high‐throughput, on‐line GCxGC‐HRMS analyses.

Conclusion

The synergy of GCxGC‐TOFMS, complementary soft ionization, and AI-based structural elucidation establishes a powerful platform for advanced metabolomics.
This approach extends compound coverage, enhances identification accuracy for unknowns, and supports broad applications in research, quality control, and biomarker discovery.

Reference

• A. Kubo et al., Mass Spectrometry, 2023, 12, A0120.
• Z. Lai, H. Tsugawa et al., Identifying metabolites by integrating metabolome databases with mass spectrometry cheminformatics, 2017.