SpecTUS: Spectral Translator for Unknown Structures Annotation from EI-MS Spectra.

Area of Science:

• Chemistry
• Computational Chemistry
• Spectroscopy

Background:

• Compound identification from mass spectra is crucial for drug discovery and forensics.
• Existing methods rely on database searches, which are limited by the size of spectral libraries and cannot identify novel compounds.

Purpose of the Study:

• To introduce SpecTUS, a deep learning model for *de novo* structural annotation of electron ionization mass spectra (EI-MS).
• To enable direct translation of EI-MS spectra into molecular structures, bypassing the need for reference databases and facilitating the discovery of unknown compounds.

Main Methods:

• Development of SpecTUS, a deep learning model utilizing gas chromatography EI-MS data.
• Training and evaluation of the model on a large dataset (NIST 20) for *de novo* structural annotation.

Main Results:

• SpecTUS achieved perfect reconstruction of 43% of compounds with a single suggestion on a held-out test set.
• The model's single suggestion showed improved Tanimoto similarity compared to hybrid database search for 76% of the test set.
• With ten suggestions, SpecTUS reached 65% perfect reconstructions, outperforming hybrid search on 84% of the test set.

Conclusions:

• SpecTUS significantly outperforms standard database search techniques for compound identification from EI-MS.
• The model's ability to identify novel compounds makes it a valuable tool for small molecule discovery and fields requiring comprehensive compound analysis.