MSNovelist: de novo structure generation from mass spectra

Area of Science:

• Computational chemistry
• Spectroscopy
• Artificial intelligence

Background:

• Current small molecule structure elucidation relies on spectral matching, limiting de novo prediction for novel compound classes.
• Tandem mass spectrometry (MS2) provides rich structural information but de novo interpretation remains challenging.

Purpose of the Study:

• To introduce MSNovelist, a novel computational tool for de novo structure generation from MS2 spectra.
• To evaluate MSNovelist's performance against existing methods and database searches.

Main Methods:

• MSNovelist integrates molecular fingerprint prediction with an encoder-decoder neural network architecture.
• The model generates candidate structures directly from MS2 spectral data without prior structural knowledge.

Main Results:

• MSNovelist achieved 25% first-rank correct structure predictions and 45% overall retrieval on a large dataset (3,863 spectra).
• The tool reproduced 61% of correct database annotations, demonstrating de novo prediction capabilities.
• Performance was validated in the CASMI 2016 challenge (26% correct, 57% retrieved) and on a bryophyte dataset, outperforming database candidates.

Conclusions:

• MSNovelist offers a powerful de novo structure prediction method for small molecules using MS2 data.
• The tool effectively complements library-based annotation, especially for novel compounds and underrepresented chemical classes.
• This approach advances the field of structure elucidation by enabling prediction beyond known spectral libraries.