Spec2Vec: Improved mass spectral similarity scoring through learning of structural relationships

Area of Science:

• Computational chemistry
• Metabolomics
• Bioinformatics

Background:

• Spectral similarity is a proxy for structural similarity in metabolomics.
• Existing methods like cosine-based scores have limitations in accurately reflecting structural similarities.
• There is a need for improved spectral similarity scoring methods in tandem mass spectrometry (MS/MS) analyses.

Purpose of the Study:

• Introduce Spec2Vec, a novel spectral similarity score.
• Improve the accuracy of structural similarity assessment in MS/MS-based metabolomics.
• Enhance scalability for large-scale database searches.

Main Methods:

• Developed Spec2Vec, inspired by Word2Vec natural language processing algorithm.
• Learned fragmental relationships from a large dataset of MS/MS spectra (GNPS libraries).
• Derived abstract spectral embeddings for similarity assessment.

Main Results:

• Spec2Vec scores show better correlation with structural similarity than traditional cosine-based scores.
• Demonstrated improved performance of Spec2Vec in library matching and molecular networking.
• Spec2Vec offers computational scalability, enabling rapid structural analogue searches.

Conclusions:

• Spec2Vec provides a more accurate and scalable method for spectral similarity assessment in metabolomics.
• The novel approach enhances the utility of MS/MS data for molecular identification and network analysis.
• Spec2Vec represents a significant advancement for computational metabolomics and drug discovery.