Learning From Repository-Scale Untargeted Metabolomics Data

Untargeted metabolomics data often face the challenge of limited annotation, with only a minority of mass spectra confidently identified using spectral libraries. Dr. Wout Bittremieux will present a transformative data-driven approach that has led to the creation of a propagated spectral library, known as the "suspect" spectral library, derived from repository-wide molecular networking results on the GNPS platform. Through the reanalysis of over 500 million mass spectra in 1335 publicly available datasets, the group compiled a novel spectral library consisting of 87,916 new reference spectra that are structurally related to known reference molecules. Dr. Bittremieux will demonstrate how the suspect library enables the discovery of novel molecules, effectively doubling the annotation rate on average. This advancement provides a powerful tool for researchers to explore previously inaccessible molecular landscapes, fostering novel biological insights. Additionally, Dr. Bittremieux will discuss recent advancements in the creation of the suspect library. Focusing on improved spectrum clustering techniques, our falcon tool minimizes data redundancy, streamlines data analysis, and enhances the interpretability of molecular networking results. Next, Dr. Bittremieux will introduce the Simba tool, a transformer neural network approach developed to predict the structural similarity between molecules based on their mass spectra. Unlike traditional modified cosine similarity, which is limited to single modifications, Simba can accurately identify molecules differing by multiple modifications. This significantly increases the number of structural analogues discoverable from untargeted mass spectrometry data. Together, these innovations mark a significant leap towards repository-scale molecular discovery, vastly amplifying the biological knowledge that can be derived from untargeted metabolomics experiments.