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

Elucidating the structure of a chemical compound is crucial in various fields. Current methods rely on matching mass spectra to spectral databases, failing for novel molecules. This paper introduces Spec2Mol, a deep learning architecture for recommending molecular structures from mass spectra. Inspired by Speech2Text models, Spec2Mol uses an encoder-decoder architecture. The encoder learns spectra embeddings, and the decoder (pre-trained on a large chemical structure dataset) reconstructs SMILES sequences. Evaluation shows Spec2Mol identifies key molecular substructures and performs comparably to existing methods, particularly for molecules absent from training databases.

Eleni E. Litsa, Vijil Chenthamarakshan, Payel Das, Lydia E. Kavraki