MASSISTANT: A deep learning model for De Novo molecular structure prediction from EI‑MS spectra via SELFIES encoding

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Artificial Intelligence

Background:

• Gas chromatography-electron impact mass spectrometry (GC-EI-MS) is crucial for identifying volatile and semi-volatile compounds.
• Manual interpretation of EI-MS spectra for unknown compounds is time-consuming and requires significant expertise.
• Existing EI-MS databases are incomplete, limiting the identification of novel molecular structures.

Purpose of the Study:

• To introduce MASSISTANT, a novel deep learning model for de novo molecular structure prediction from EI-MS spectra.
• To leverage SELFIES encoding for direct molecular structure generation.
• To provide mass spectrometry scientists with an automated tool for enhanced compound identification.

Main Methods:

• Development of a deep learning model named MASSISTANT.
• Utilizing SELFIES encoding to represent molecular structures.
• Training the model on low-resolution EI-MS spectra of compounds under 600 Da.
• Evaluating performance on curated and broad datasets, including the NIST database.

Main Results:

• MASSISTANT directly predicts molecular structures from EI-MS spectra.
• Performance varied with dataset curation, achieving up to 54% exact predictions (Tanimoto score = 1) on a focused dataset, compared to 10% on the full NIST dataset.
• The model demonstrates the capability of deep neural networks in interpreting complex fragmentation patterns.

Conclusions:

• Deep learning models like MASSISTANT can effectively generate chemically valid molecular structures from EI-MS data.
• MASSISTANT offers a powerful tool to assist scientists in interpreting mass spectrometry analyses.
• This approach enhances the elucidation of molecular structures, substructures, and functional groups in GC-EI-MS.