Topic modeling for untargeted substructure exploration in metabolomics

Significance Tandem MS is a technique for compound identification in untargeted metabolomics experiments. Because of a lack of reference spectra, most molecules cannot be identified, and many spectra cannot be used. We present MS2LDA, an unsupervised method (inspired by text-mining) that extracts common patterns of mass fragments and neutral losses—Mass2Motifs—from collections of fragmentation spectra. Structurally characterized Mass2Motifs can be used to annotate molecules for which no reference spectra exist and expose biochemical relationships between molecules. For four beer extracts, without training data, we show that, with 30 structurally characterized Mass2Motifs, we can annotate approximately three times as many molecules as with library matching. These Mass2Motifs were validated in reference spectra from Global Natural Products Social Molecular Networking (GNPS) and MassBank. The potential of untargeted metabolomics to answer important questions across the life sciences is hindered because of a paucity of computational tools that enable extraction of key biochemically relevant information. Available tools focus on using mass spectrometry fragmentation spectra to identify molecules whose behavior suggests they are relevant to the system under study. Unfortunately, fragmentation spectra cannot identify molecules in isolation but require authentic standards or databases of known fragmented molecules. Fragmentation spectra are, however, replete with information pertaining to the biochemical processes present, much of which is currently neglected. Here, we present an analytical workflow that exploits all fragmentation data from a given experiment to extract biochemically relevant features in an unsupervised manner. We demonstrate that an algorithm originally used for text mining, latent Dirichlet allocation, can be adapted to handle metabolomics datasets. Our approach extracts biochemically relevant molecular substructures (“Mass2Motifs”) from spectra as sets of co-occurring molecular fragments and neutral losses. The analysis allows us to isolate molecular substructures, whose presence allows molecules to be grouped based on shared substructures regardless of classical spectral similarity. These substructures, in turn, support putative de novo structural annotation of molecules. Combining this spectral connectivity to orthogonal correlations (e.g., common abundance changes under system perturbation) significantly enhances our ability to provide mechanistic explanations for biological behavior.