MassGenie: A Transformer-Based Deep Learning Method for Identifying Small Molecules from Their Mass Spectra.

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in chemistry

Background:

• The inverse problem of mass spectrometric molecular identification remains largely unsolved, particularly in metabolomics, due to limited experimental spectra.
• Predicting molecular structure from mass spectra is crucial for identifying unknown compounds.

Purpose of the Study:

• To develop a novel method, MassGenie, for predicting molecular structures from mass spectra using deep neural networks.
• To address the challenge of de novo small molecule structure prediction from experimental mass spectra.

Main Methods:

• Utilized a transformer-based deep neural network trained on approximately 6 million in silico generated chemical structures and their corresponding molecular fragments.
• Treated mass spectral interpretation as a language translation problem, with mass spectra as the source and molecular structures (SMILES) as the target.
• Incorporated VAE-Sim for generating similar candidate molecules to the top predicted structure.

Main Results:

• MassGenie successfully learned effective properties of mass spectral fragmentation and valency space without explicit rules.
• Achieved 53% precise correct identification (49/93) in the CASMI challenge for molecules under 500Da.
• Demonstrated high effectiveness on both in silico generated and experimentally obtained mass spectra, acting as a Las Vegas algorithm.

Conclusions:

• Transformer deep neural networks are highly effective for mass spectral interpretation and de novo small molecule structure prediction.
• MassGenie opens new avenues for identifying unknown compounds by generating candidate structures not limited to existing libraries.
• The method shows significant promise for advancing metabolomics and drug discovery through accurate structure elucidation.