Metabolomics workbench: data-driven classification and correlation analysis between metabolite classes, biomarkers, and drugs.

Public repositories now accumulate an unprecedented amount of metabolomics data. Despite the opportunities for large-scale meta-analyses opened by these repositories, the deposited data rarely becomes systematically re-analyzed. We conducted a systematic metadata-based re-analysis of all human blood metabolomics studies in the Metabolomics Workbench performed using Orbitrap or QTOF mass spectrometry, encompassing 44 165 samples. After harmonizing metabolites with PubChem identifiers and embedding them using SMILES representations, we generated a map of 6752 blood metabolites forming 25 structural clusters, and subsequently performed correlation analysis between these clusters, biomarkers, and drugs. Analysis of 337 120 metabolite pairs detected in at least 10 studies showed that structurally related clusters display correlated abundances (ρ up to 0.6). Correlation patterns among established clinical biomarkers revealed candidate metabolic correlates, several with ρ > 0.7. We also developed open-source software enabling similar analyses for any metabolite across the Workbench data explored. Additionally, curating drug-containing datasets enabled meta-analyses for compounds such as acetaminophen, metoprolol, and atenolol, identifying both known and novel metabolic signatures. Besides the biological insights, we have found that Orbitrap and QTOF tend to detect different metabolites, but they give quantitatively coherent results on overlapping compounds. Overall, the present study provides the methodology and software to perform large-scale mining of public metabolomics data enabling its re-analysis to obtain novel biomedical insights.