MIST

Summary

MIST (Metabolite Inference with Spectrum Transformers) is a machine learning tool designed to predict molecular fingerprints directly from tandem mass spectrometry (MS/MS) data. It employs a transformer-based neural network architecture that maps mass spectrometry peaks (m/z and intensity) to chemical formulas and neutral losses, converting spectra into continuous embeddings to enable search and annotation against molecular databases.

Methodology

MIST models the relationship between spectra and chemical structures using: - Formula Transformers: An architecture that explicitly incorporates chemical formula relationships to map peaks to candidate chemical formulas, making peaks more "word-like". - Fingerprint Prediction: Predicts binary molecular fingerprints (such as Morgan fingerprints) from the spectral representation. - Database Lookup: Queries predicted fingerprints against chemical databases like PubChem to rank potential molecular structures.

Performance and Evaluation

According to a benchmark analysis by Khoo & Barzilay (2026), MIST shows: - Random Split Performance: Achieves a Jaccard score of 0.547 on the NPLIB1 dataset, 0.674 on MassSpecGym, and 0.622 on NIST2023. - Scaffold Split Performance: Drops significantly to a Jaccard score of 0.241 on NPLIB1, 0.267 on MassSpecGym, and 0.255 on NIST2023. - Generalization: Demonstrates limited capacity to generalize to unseen chemical scaffolds compared to foundation models like DreaMS or simple nearest-neighbour baselines.

Citations

• Goldman, S., et al. (2023). Formula Transformers for Molecular Structure Elucidation. Nature Machine Intelligence, 5, 965–979. DOI: 10.1038/s42256-023-00717-3
• Khoo, L.M.S., & Barzilay, R. (2026). Why machine learning fails at mass spectrometry for small molecules. Nature Metabolism, 8, 1247–1249. DOI: 10.1038/s42255-026-01544-6. Source paper: s42255-026-01544-6.pdf