Topic modeling for untargeted substructure exploration in metabolomics.

Area of Science:

• Life Sciences
• Biochemistry
• Computational Biology

Background:

• Untargeted metabolomics holds great potential for life sciences research.
• Current computational tools are limited in extracting biochemically relevant information from mass spectrometry data.
• Existing methods rely on spectral fragmentation patterns, requiring standards or databases for molecule identification.

Purpose of the Study:

• To develop a novel analytical workflow for extracting biochemically relevant features from untargeted metabolomics data.
• To address the limitations of existing tools by exploiting all available fragmentation data in an unsupervised manner.
• To enable de novo structural annotation and enhance mechanistic explanations in biological systems.

Main Methods:

• An unsupervised analytical workflow was developed to process all fragmentation data from metabolomics experiments.
• Latent Dirichlet Allocation, a text-mining algorithm, was adapted for metabolomics datasets.
• The approach identifies "Mass2Motifs" – biochemically relevant molecular substructures represented by co-occurring fragments and neutral losses.

Main Results:

• The workflow successfully extracts molecular substructures (Mass2Motifs) from spectral data.
• Molecules can be grouped based on shared substructures, independent of classical spectral similarity.
• This spectral connectivity, combined with orthogonal correlations, improves the ability to explain biological behavior mechanistically.

Conclusions:

• The developed workflow enhances the utility of untargeted metabolomics by extracting key biochemical information.
• Mass2Motifs provide a novel way to group and annotate molecules based on shared biochemical features.
• This approach significantly advances the potential of metabolomics for mechanistic biological discovery.