An end-to-end deep learning framework for translating mass spectra to de-novo molecules

Area of Science:

• Computational Chemistry
• Cheminformatics
• Artificial Intelligence in Chemistry

Background:

• Elucidating chemical structures is crucial for drug discovery, precision medicine, and biomarker identification.
• Current methods rely on mass spectra and spectral databases, failing for novel molecules.
• There is a need for computational tools to predict molecular structures from mass spectral data.

Purpose of the Study:

• To propose Spec2Mol, a deep learning architecture for molecular structure recommendation using only mass spectra.
• To address the limitations of existing methods for novel chemical compounds.
• To enable structure elucidation for molecules absent in reference spectral databases.

Main Methods:

• Developed Spec2Mol, an encoder-decoder deep learning architecture inspired by Speech2Text models.
• The encoder generates embeddings from mass spectra.
• The decoder, pre-trained on diverse chemical structures, reconstructs SMILES (Simplified Molecular Input Line Entry System) sequences.

Main Results:

• Spec2Mol successfully identifies key molecular substructures directly from mass spectra.
• Evaluated performance based on molecular similarity between recommended and original structures.
• Achieved comparable performance to existing fragmentation tree methods, especially for novel or uncatalogued compounds.

Conclusions:

• Spec2Mol offers a robust deep learning solution for molecular structure recommendation from mass spectrometry.
• The model demonstrates significant potential in accelerating drug discovery and related fields.
• Spec2Mol provides a valuable tool for analyzing novel chemical entities where reference data is scarce.